improvise/context/test-evidence/command.md

# `improvise-command` — test evidence

Why each of these tests exists, what bug it reproduced, and what fault was
injected to prove it can fail. Keyed by **exact function name**; a rename
orphans the entry, and `scripts/check-comment-runs.py --check-evidence` reports
that (improvise-9c8c, design-principles §7).

Entries cover tests and the helpers whose *shape* is the evidence.

---

## `ui/reduce.rs`

### a_view_slice_command_does_not_move_the_authoritative_view_state

This test used to assert the opposite, under the name
`view_slice_commands_land_in_the_authoritative_view_state`, and it passed against
a path no host walks (improvise-s17o): `apply_command`'s fallback arm folded a
wire view command into `app.view_state`, the *authoritative* slice. Nothing in
production reached it — `Sessions::apply_checked` folds a view command into the
asking client's mirror — so the behaviour it pinned was wrong and safe only
because it was dead. The model is shared, the UI is not (principle 5, gotcha 36);
a client's `SetStatus` must not move the host's status line.

The fixture writes `"host"` into the status line first, so the assertion
distinguishes "the fold was removed" from "the field happened to be empty" — with
the default empty string, deleting the fold and leaving it in place look the same
for `SetStatus`. Restoring `reduce_view(&mut app.view_state, other)` fails on the
first field.

#### Helper — `checked`

Not a test — the fixture every `reduce.rs` test below reaches the door through.

For the tests whose subject is not staleness: a command the current state
accepts, which every command built by `caught_up` is by construction. The tests
whose subject *is* staleness call `Checked::against` themselves and assert on the
`Rejection`, so neither kind can pass by accident through the other's
convenience.

### switch_view_moves_the_client_and_leaves_the_documents_default

The successor to `switch_view_moves_the_document_and_the_client_together`, which
asserted the pre-improvise-jytm lockstep. Which view a client is on is now that
client's; `Workbook::active_view` is the *default* a fresh client opens and what
`.improv` writes. The fault it rules out is the old behaviour returning by inertia
— any writer that moves the workbook's copy from a client's switch fails it.

Fault injected 2026-09-20: restoring `App::switch_to_view`'s
`workbook.switch_view(name)` call fails this and
`the_documents_default_view_outlives_this_clients_switches`.

### a_wire_view_switch_records_history_the_way_a_key_switch_does

A wire-driven view switch owes the same history a key-driven one leaves behind.

`effect::ViewBack` pops `view_state.view_back_stack`, and the only thing
that ever pushes onto it is `effect::SwitchView`. While `apply_command` performed
the switch itself — calling `App::switch_to_view` directly rather than running
the effect — the stack stayed empty, so a client that switched over the wire
could not go back, though the same two keypresses could. Wave 4 is what made that
reachable: before `ModelCommand` there was no wire way to ask.

Nothing else in the suite could see it.
`switch_view_moves_the_client_and_leaves_the_documents_default` checks where a
switch *landed*; this checks what it *recorded*.

Bug demonstrated 2026-09-19 (Copilot, PR #11): the `ViewBack` left the document
on "Second". Since improvise-jytm the step back is driven through `effect::ViewBack`
rather than `ModelCommand::ViewBack`, which left the wire with that ticket — the
stack it pops is one client's history, so a nullary wire command could not address
it (gotcha 38). What the test asserts is unchanged.

### switching_views_projects_the_axis_widgets_only

Switching views changes what the axis-bearing widgets read, so they are told; the
list of views itself did not change, so it is not.

### an_inbound_projection_is_ignored

Projections travel one way. A client that sends one back must not be able to
write it into the server's state.

### a_keypress_that_changes_the_model_projects_it

The key path is the model's door, and it reports the widgets it touched so a host
can project them.

### a_view_only_keypress_projects_nothing

A keypress that moves only this client's cursor reaches no other client
(gotcha 3).

### every_model_command_names_an_effect_that_declares_a_model_change

The derivation, stated as a check. This is what "derived, not invented" means
operationally (`doctrine(derive_the_enum_from_the_effects)`). A variant pointed
at a `ViewOnly` effect would be a wire message that changed the document for the
client that sent it and for nobody else — `Dirty::of` would report `NOTHING`, and
no projection would ever correct the others (gotcha 35).

#### Helper — `model_app`

Not a test — the fixture the three `every_model_command` walks run against.

`app()` with a cell actually stored, so a surrogate for it resolves through the
index rather than only through the layout. The walks over `every_model_command`
run against this: a `SetCell` whose record cannot be resolved produces *no
effects*, and every one of those tests asserts on the effects, so the fixture has
to name something real or they would all fail for a reason that is not theirs.

### no_model_command_commits_another_clients_session_local_state

improvise-3x3u: the wire vocabulary is the *addressable* model effects, not all
of them.

`Sessions::apply` routes a `Command::Model` with no `SessionId`, which is right
exactly when the change is addressed to no one session (gotcha 36).
`ApplyAndClearDrill` and `AcceptWizardModel` are addressed to precisely one, and
it is not the sender: an unaddressed `ApplyAndClearDrill` commits another
client's private typing (improvise-tom1) and an unaddressed `AcceptWizardModel`
swaps the document out of a dialogue the sender never opened. Both were
demonstrated against `Sessions::apply` before the variants were removed.

Re-adding either variant means making it *addressed* — carrying a `SessionId`,
with the staging map moved onto the session — not restoring the nullary form this
test forbids.

`SESSION_LOCAL_EFFECTS` is the list of the two effect names.

### the_session_local_effects_are_real_model_effects

The guard above is only worth having while the effects it names still exist and
are still model changes — otherwise it would pass by naming nothing
(`doctrine(check_the_checker)`).

### the_effect_name_reader_reads_all_three_debug_shapes

The reader above is itself a check, so it gets the injection too
(`doctrine(check_the_checker)`): one effect of each `Debug` shape, because the
leading identifier is followed by a different delimiter in each — nothing, `(`,
and ` {`.

#### Helper — `effect_type_name`

Not a test — the reader that test injects against, used by the two guards above.

The type name of an effect, read off the leading identifier of its derived
`Debug`. `Effect` is a trait object, so `std::any::type_name` is not reachable
through it; every effect derives `Debug`, and a derived `Debug` opens with the
type's own name for a unit struct, a tuple struct and a braced struct alike. So
the name is already on the value — this only reads it back.

### every_model_command_names_the_effect_of_the_same_name

The mapping is name-for-name by construction, so state that: every
`ModelCommand` variant names the effect that carries its own name.

This is the check that separates an effect from its inverse.
`every_model_command_names_an_effect_that_declares_a_model_change` cannot:
`HideItem` and `ShowItem`, `AddItem` and `RemoveItem`, `CreateView` and
`DeleteView`, `SetCell` and `ClearCell` are each two effects of the *same*
`EffectKind`, so a swapped pair clears that bar while inverting what the wire
message does. And `a_model_command_lands_where_its_keypress_lands` cannot reach
them either — most variants have no single keypress to compare against, so it
drives three of twenty-eight.

Proven by injection 2026-09-19: `HideItem` → `effect::ShowItem`, `CreateView` →
`effect::DeleteView` and `AddItem` → `effect::RemoveItem` all left
`cargo test --workspace` at 918 passing, 0 failing. This test fails on each.

The name-for-name rule is a design constraint, not a coincidence, and it is what
makes the derivation checkable at all
(`doctrine(derive_the_enum_from_the_effects)`): a variant whose effect must be
named something else is a variant that stopped being derived, and it should fail
here and be argued about rather than pass quietly.

### the_enumeration_names_each_variant_once

The `every_model_command` list covers the enum once. A duplicated entry would
leave a variant untested while the count still looked right.

#### Helper — `every_model_command`

Not a test — the enumeration that test, and the three walks above it, cover.

One value per `ModelCommand` variant. The enum is the derived set, so a test that
walks "the set" has to walk all of it. `variant_name` is the completeness guard:
it matches exhaustively, so a variant added to the protocol fails the build here
until it is listed in this vector as well.

### a_model_command_lands_where_its_keypress_lands

The point of the whole ticket: a model `Command` applied authoritatively leaves
the document exactly where the keypress bound to the same mutation leaves it.
That equivalence is what lets the two entry points share one door.

The `Dirty` is compared too, because a host with connected clients projects off
it: a command that produced the right document and the wrong `Dirty` would leave
every other client stale.

Its `pivoted()` fixture has **two** categories on different axes, because the
one-category fixture cannot tell a transpose from a cycle: with only `Region` on
Row, swapping the axes and rotating Row→Column land on the same view, and an
equivalence test that cannot distinguish the wrong effect from the right one
proves nothing (`doctrine(check_the_checker)`).

### a_stale_wire_model_command_cannot_be_applied

A wire `Command::Model` cannot reach the document without its basis being checked
(ADR 001, improvise-ondj).

Bug demonstrated 2026-09-19: `reduce_full` and `apply_command` — the functions
this module's own doc points every future host at, the browser worker included —
took a `&Command` and matched `Command::Model { cmd, .. }`, so a batch composed
against a document that had since moved landed anyway. The `..` discarded the
basis with no compile error.

**The regression guard is the signature, not this test.** Both now take a
`Checked`, whose only constructor runs `stale_basis`, so the version of this test
that reproduced the bug — `reduce_full(&mut app, &wire)` with a stale `wire` — no
longer compiles, and `reduce_full` itself no longer exists (improvise-s17o: having
no caller was the condition that let the defect hide in the first place). What is left to assert is that the constructor refuses, and that
refusing leaves the document alone: there is no half-applied batch, because the
type is obtained before anything is applied.

### a_model_commands_dirty_comes_from_its_effects_kind

The `Dirty` a model command reports is the one its effect declares, not a second
table kept alongside the enum (gotcha 35).

### a_wire_model_command_marks_the_document_dirty

improvise-gv9r. **The bug**: a client's edit left the authoritative document
reporting *clean*, on every transport.

`MarkDirty` used to be a separate effect that the *command* layer composed
alongside the mutation, and improvise-kmgs derived `ModelCommand` from the
effects **one at a time** — so the wire got `SetCell` without its companion. A
derivation correct per element and wrong for a set. "Save if dirty" was therefore
a trigger that never fired, and a host built on it loses every client edit
silently; the browser worker found that out the hard way and had to
over-approximate off the command variant instead.

The fix is the unrepresentable one: dirtiness is raised by `App::apply_effects`
off the same `EffectKind` the `Dirty` and the two basis tiers are read from
(gotcha 40), so no path can compose a document change without it. There is
nothing left to forget, and `MarkDirty` is gone rather than left as a second
answer.

The three levels are asserted separately because they are three different
effects: an effect that is only ever reached by one of them would otherwise pass
on a sibling's evidence. The negative arm — that a `ViewOnly` effect leaves the
document clean — is `a_view_only_batch_leaves_the_document_clean` in `ui/app.rs`,
and it has to live there: a view-slice `Command` never reaches `apply_effects` at
all (`apply_command` folds it straight into `ViewState`), so a negative written
against this door would pass however unconditional the raise was.

### a_wire_view_switch_leaves_the_document_clean

The inversion of `a_wire_view_switch_marks_the_document_dirty`, which asserted the
pre-improvise-jytm behaviour. A switch used to move `Workbook::active_view`, which
persists as the `Initial View` line, so it really did put the document out of step
with the file. It no longer moves any document state, so there is nothing to
persist and the browser worker's persist trigger (improvise-gv9r) must not fire.

The direction matters: the worker asks the document's `dirty` flag rather than a
table of variants, so a switch that dirtied spuriously would cost an OPFS rewrite
of the whole document per keystroke (gotcha 41). This is the negative arm — the
positive one, that a real model command *does* dirty, is the loop above it.

The consequence to keep in view is a product one, filed as improvise-azhb rather
than decided here: a TUI user who switches view and saves now writes the
`Initial View` the document was loaded with.

---

## `ui/effect.rs`

### the_document_surface_runs_on_a_non_filesystem_backend

improvise-fuj: every effect that touches a document goes through `App::storage`,
so the whole save/load/import/export surface works on a backend that is not a
filesystem. This is what a browser worker supplies (improvise-i34); here
`MemoryFs` stands in for it, and the test would fail if any of these effects
reached for `std::fs`.

### view_only_effects_leave_shared_state_alone

improvise-mae: a declared `kind()` is only worth having if it matches what the
effect actually does. Rather than restate each declaration, this applies the
effect and checks the shared state it claims not to touch is genuinely unchanged
— a `ViewOnly` effect that quietly wrote a cell would never reach connected
clients.

No exemptions: since improvise-8uz moved the cursor out of `View`,
`SetSelected`/`SetRowOffset`/`SetColOffset` touch nothing shared, so every effect
in the list must clear the same bar.

### structural_effects_declare_their_kind

The structural effects, each pinned to the kind it must declare.

The bug this guards: `AddCategory` and `RemoveCategory` declared `CellData`, but
both go through `Workbook`, whose `on_category_added` / `on_category_removed`
rewrite **every** view's axis map. That is a re-layout, so a client told only
"cell data changed" would keep rendering the old grid shape. The two "actually
change" tests beside it cannot see this — a category change moves model data
*and* view config, so it passes either list. Only pinning the declaration does.

### add_category_past_the_limit_reports_the_failure

Bug (improvise-djn): `AddCategory` discarded the `Result` from
`Workbook::add_category`, so pressing `n` in the category panel with the
12-regular-category limit already reached reported the optimistic
`Added category "X"` its command had queued and added nothing. The effect runs
last precisely so its failure can overwrite that message (improvise-6t3) — it
just never wrote one.

### add_formula_adds_target_item_to_category

Regression: `AddFormula` must add the target item to the target category so the
grid layout includes cells for it. Without this, a formula like
`Margin = Profit / Revenue` would be registered but invisible in the grid.

### abort_chain_short_circuits_apply_effects

`AbortChain` must cause subsequent effects in the same `apply_effects` batch to
be skipped, and the flag must reset so the next dispatch starts clean.

### save_without_file_path_aborts_the_chain

BUG (improvise-8tk): `Save` reported failure only through the status bar, so the
`Quit` that `wq`/`ZZ` chains after it landed anyway and the whole session was
discarded without the user seeing the error. A failed save must abort the rest of
the batch.

### failed_save_as_aborts_the_chain

BUG (improvise-lae, found reviewing it): improvise-8tk fixed **one arm** of a rule
with two. `Save` aborts the batch when the write fails; `SaveAs` reported the
error to the status bar and let the batch continue, which is the same silent
discard 8tk closed — the shape
`doctrine(one_arm_of_a_symmetric_rule_is_evidence_about_that_arm_only)` names,
and the three tests above it are evidence about `Save` only.

It is reachable today: the `:` line composes with ` . `, so `:w /bad/path . q`
builds exactly this batch, and `ExecuteCommand` concatenates the two commands'
effects into one. improvise-lae's save prompt also commits to `SaveAs`, so the
arm was about to carry more traffic.

Note what the abort does and does not buy the prompt: the typed path survives for
the user to correct, but the message does not reach the screen, because
`draw_bottom_bar` draws the minibuffer instead of the status line while a prompt
is open. That is pre-existing and reported separately; the data-loss half is what
this fixes.

### successful_save_as_lets_the_chain_continue

The positive arm: a `SaveAs` that succeeds must not abort, or every `:w <path>`
would swallow whatever followed it. Without this, "always abort" passes
`failed_save_as_aborts_the_chain`.

### enter_edit_at_cursor_uses_target_mode_not_app_mode

`EnterEditAtCursor` must use its `target_mode` field, *not* whatever
`app.view_state.mode` happens to be when applied. The previous implementation
branched on `app.view_state.mode.is_records()`; the parameterized version trusts
the caller (keymap or composing command).

### enter_edit_at_cursor_reads_post_batch_layout

improvise-ehz: the edit buffer must be loaded from layout state that already
reflects earlier effects in the same batch. `apply_effects` only rebuilds *after*
the batch, so a preceding `RebuildLayout` is what makes the cursor resolve to the
post-mutation cell. Removing the row-0 item slides a different cell under the
cursor; the buffer must show that cell's value, not the pre-batch one.

### set_buffer_does_not_special_case_the_search_name

`SetBuffer` no longer treats "search" as an alias for the search query
(improvise-gdl): it is a plain buffer write whatever the name says, and the query
is reached through `SetSearchQuery`. Commands pick between the two in
`cmd::core::write_buffer`.

### apply_and_clear_drill_coord_rename_moves_every_measure_in_the_row

improvise-c5v: in wide format a records row groups one cell per measure, so
renaming a coordinate has to move every one of them. The long-format code moved
only the cell under the cursor, which would now split the row in half.

### import_headless_csv_short_rows_surface_warning_in_status

BUG (improvise-k8i): `ImportJsonHeadless` parsed CSVs with `parse_csv`, which
sends short-row warnings to stderr — invisible inside the TUI. Short rows must
surface in the status message, naming the affected column, never silently
succeed.

#### Note — the drill tests drive `apply_effects`, not `apply`

Several `ApplyAndClearDrill` tests call `app.apply_effects(vec![...])` rather
than `effect.apply(&mut app)`. That is deliberate: the dirty flag is raised off
the `EffectKind` now (improvise-gv9r, gotcha 45), so driving the effect directly
would assert a flag nothing in production sets.

### every_file_writing_effect_aborts_the_batch_on_failure

improvise-z05r, the **third** instance of one rule: improvise-8tk fixed `Save` and
not `SaveAs`; PR #27 found `SaveAs`; `ExportCsv` was the arm still missing, so a
`Binding::Sequence` or a `:` line composing an export with anything after it
continued past a failed write.

The fix is structural rather than a third patch: all three now report through one
`report_file_write(app, outcome)`, which writes the status message either way and
sets `abort_effects` on `Err`. That is not sufficient by itself — a fourth writer
could still be written without calling it — so the test names the three, and a
fourth has to come here and be named. Naming is the mechanism; the shared function
is what makes the naming cheap to honour.

Each case drives `App::apply_effects` with `[writer, ChangeMode(Quit)]` against a
path under a directory that does not exist, and asserts both `last_batch_aborted()`
and that the trailing mode change never landed.

Injected fault: drop `app.abort_effects = true` from `report_file_write`. Fails by
name, along with the four older abort tests — which is the point: they and this one
now share one mechanism.

Injected fault, narrower: route `ExportCsv` around the reporter and write its status
directly. Only this test fails, naming `ExportCsv`, with 460 passing — the exact
state the ticket describes.

### a_batch_that_writes_nothing_leaves_the_document_clean

improvise-748f. `ViewNavigate { forward: false }` emits `ApplyAndClearDrill`
unconditionally, and that effect returns immediately when `drill_state` is `None`.
It declares `Model(CellData)` regardless — `Effect::kind` takes no `App`, so the
classification cannot be state-dependent — so `apply_effects` raised
`model_state.dirty` on a document nothing had touched. Press view-back, press `:q`,
and the quit is refused; autosave arms thirty seconds later.

The older test here, `start_drill_and_apply_clear_drill_with_no_edits`, pinned the
wrong behaviour in an assertion message arguing it was the same over-approximation a
failed `LoadModel` gets. It is not: a succeeding `LoadModel` leaves a document that
really does differ from the file its path names, and that remains declared. This was
a wrong answer to no divergence at all.

### a_cell_data_effect_no_raw_cell_records_still_dirties_the_document

The guard against the cure improvise-748f **proposed**, which is unsound. The ticket
suggested raising `dirty` off `bump_changed_records`' per-record fingerprint of the
raw cells, on the reading that `CellData` means cells. Five of the ten `CellData`
effects change the document without writing a raw cell: `AddItem`, `AddItemInGroup`
and `SortData` (which reorders — and `.improv` carries the order), and `AddFormula`
and `RemoveFormula`, whose results live in `Model::formula_cache` and not in the
store. Observing raw cells would have left every one of them reading clean, which is
a *lost edit* — the opposite direction from the spurious prompt it was fixing.

So the observation is `document_content` — `document_config` plus the cells in
iteration order and the formulas — and **only the `CellData` arm reads it**. The
`LayoutGeometry` arm keeps the narrow `document_config`. A first version pointed both
arms at the wide fingerprint, and review benchmarked what that costs at 18,000 cells
in release over 200 batches after warm-up: one `SetCell` batch 8.07 ms → 10.21 ms
(+27%, a constant factor on a path that was already O(cells) through
`record_fingerprints`), but one `SetNumberFormat` batch **78 µs → 2.63 ms, 33×** —
an *asymptotic* change from O(config) to O(cells + config) on the held-key path that
`[`/`]`, axis cycle, hide/show, group toggle, transpose and prune toggle all take.
~380/sec where they had been ~12,800/sec. The split costs no correctness because all
five blind effects are `CellData`; that every `LayoutGeometry` effect really is
visible to the narrow fingerprint is asserted separately, and not left as a reading.

`WholeWorkbook` stays *declared*, and that is the remaining over-report: a `LoadModel`
whose incoming content happens to be identical still marks the document dirty, and a
failed one does too. Both cost one redundant write.

The `SortData` case needs two cells inserted out of sorted order, or it reorders
nothing and its case asserts the opposite of what it names. The `AddFormula` case
removes the fixture's formula first, or adding it is a no-op.

Injected fault: narrow `document_content` back to the configuration it hashed before.
Fails naming `SortData`, with five other dirty-flag tests.

### every_layout_geometry_effect_is_visible_to_the_narrow_fingerprint

The guard that makes the two-fingerprint split safe rather than a trap. Splitting
`CellData` (wide) from `LayoutGeometry` (narrow) is only sound while every effect
wearing `LayoutGeometry` writes something `document_config` hashes; an effect it
cannot see would leave the document reading clean **forever**, which is the same
lost-edit failure the wide fingerprint exists to prevent, arriving by the other door.

The thirteen are enumerated by brace-matching the `impl Effect` blocks, not by
grepping — the naive grep reports fourteen, picking up `HelpPageSet`, which is
`ViewOnly` and is simply the last `pub struct` before `#[cfg(test)]` (gotcha 38 names
this exact artefact). Each case asserts its effect really declares
`Model(LayoutGeometry)` before asserting anything else, so a reclassified effect fails
here rather than silently leaving the list unrepresentative.

**Four** cases carry a `prepare` step run **directly against the workbook** and not
through an effect, because the interesting state is the precondition. What makes such
a case honest is the **no-op-fails** property: without its precondition the effect
writes nothing, an unwritten document leaves the fingerprint unmoved, and the case
therefore *fails* rather than passing vacuously.

Measured, by deleting each `prepare` in turn, **two** of the four are load-bearing in
exactly that way and two are not:

- `DeleteView` needs a second view; without it, `dirty` reads false and the case fails.
- `ShowItem` needs a hidden item; without it, `dirty` reads false and the case fails.
- `ToggleGroup` is **not** load-bearing here. `ToggleGroup::apply` calls
  `toggle_group_collapse` unconditionally — its membership check gates only
  `reset_cursor` — so `collapsed_groups` moves whether or not the group is real, and
  the case passes with the `prepare` deleted. The step documents the realistic input;
  it is the *cursor* arm, tested elsewhere, that needs a real group.
- `TransposeAxes` is **not** load-bearing either, and its step is literally a no-op:
  `View::on_category_added` already assigns the first regular category `Row` and the
  second `Column`, so the fixture arrives with `Type`/`Month` differing. The step
  states the precondition rather than establishing it, and the case passes without it.

`dirty` is cleared **after** `prepare` and before the effect. That clear is defensive
hygiene for a *future* `prepare` that goes through `apply_effects`, and it is not what
makes any case honest today: `dirty` is raised **only** inside `App::apply_effects`,
and no `prepare` in this test runs an effect, so no preparation can set it. Demonstrated
by removing the clear **and** blinding `document_config` to `0` together: all thirteen
cases still read `dirty == false`. An earlier version of this entry claimed the
opposite — that without the clear "all four cases pass against any fingerprint" — and
measurement contradicts it.

Injected fault: `document_config` returns `0`. Fails, with five other dirty-flag tests
across three files. Injected fault, narrower and the one that proves the test is
pointed at the right thing: drop `view.number_format` from `document_config`. Fails
naming **`SetNumberFormat`**, with only one other test failing anywhere.


---

## `ui/app.rs`

### model_state_and_view_state_are_constructible

improvise-3vr: `ModelState` and `ViewState` are the named slices of `App` state
introduced by the vb4 refactor. Step 1 only requires that the types exist and are
constructible; subsequent steps move fields in.

#### Note — the `ViewState: Send` static assertion

improvise-h0b: `ViewState` is the portable slice — the one thing serialized into
the Snapshot a newly-connected client receives (improvise-q08). `Send` is the
cheapest compile-time proxy for "portable": a field that cannot cross a thread
boundary certainly cannot cross a serialization one. Before h0b this failed to
compile, because `drill_state: Option<DrillState>` holds an `Rc<Vec<RecordRow>>`.
Session-local state lives on `App` instead. It is a `const _: fn()` in the test
module, not a named test, so it has no heading of its own.

### app_view_state_owns_ui_session_fields

improvise-ew0: `ViewState` owns the UI session slice — mode, status, search,
panels, navigation, drill, yanked, buffers. `App` accesses them through
`view_state`.

### view_state_survives_a_json_round_trip

improvise-a2x: `ViewState` is what a newly-connected client receives as a Snapshot
(improvise-q08), so it has to survive a round trip through the wire format.
Populated with non-default values in every field family — a mode carrying a
`MinibufferConfig`, a navigation frame, a yanked cell, a keyed buffer, an armed
prefix — because a roundtrip of `ViewState::default()` would pass while proving
nothing.

### the_documents_default_view_outlives_this_clients_switches

The replacement for `the_two_active_view_names_never_diverge`, and it is a
replacement rather than a deletion on purpose: improvise-jytm makes the two copies
diverge deliberately, and a deliberate split with no guard is how one silently
regresses into the other.

What it pins, all four of which have a silent failure mode. A switch moves the
client's copy and **not** the document's default. `App::view()` — what
`rebuild_layout` and `cmd_context` read — follows the client, so the split is
observable in the layout and not merely in a string. A switch to a view the
document does not have moves nothing, because the slice keys `ViewState::cursors`
and `grid_in` indexes the workbook by it. And `install_workbook` still seats this
client on the incoming document's default and clears the cursor map, since the
outgoing document's view names mean nothing in the new one (gotcha 24).

Fault injected 2026-09-20: `App::switch_to_view` calling
`workbook.switch_view(name)` again fails the second assertion.

### moving_the_cursor_does_not_touch_the_shared_workbook

improvise-8uz: moving the cursor must leave the shared document untouched, so that
one client scrolling cannot move another's viewport. Before the move these three
effects wrote into `model_state.workbook`, and this assertion would have failed.

### app_owns_the_session_local_fields_not_view_state

improvise-h0b: the two session-local fields moved off `ViewState` and onto `App`.
Each fails both membership tests: `ImportWizard` is an `improvise-io` type (not
portable) and a half-open dialog (not snapshot-safe); `DrillState` holds an `Rc`
(not portable — see the `Send` assertion) and stages uncommitted edits (not
snapshot-safe).

### staged_drill_edits_are_unreachable_from_view_state

improvise-h0b: a fresh client must not inherit another session's staged,
uncommitted drill edits. Those edits are unreachable from `ViewState`, so a
Snapshot built from it cannot carry them.

### switching_views_restores_each_views_own_cursor

improvise-8uz: the cursor is **per view**. Written as a characterization test
against the old `View`-owned fields and kept verbatim across the move to
`ViewState::cursors`, because a refactor about *where* state lives must not change
*what the app does*.

### the_improv_format_does_not_persist_the_cursor

improvise-8uz: the ticket asserted these fields were "PERSISTED in .improv". They
were not — the grammar has no cursor production and the writer emitted none. That
is what made the move free: there was no restore-where-you-left-off behaviour to
lose, and no format change to make. The assertion still guards the format going
forward.

### entering_records_mode_sorts_existing_data

Pressing `R` to enter records mode sorts existing data by `CellKey`, so display
order is deterministic regardless of insertion order.

*Note:* this test's doc comment carried a stray first line from a neighbouring
test ("Regression: pressing `o` in an empty records view should create the") that
was never finished. It was dropped in the comment-cap wave; the sentence above is
what the test actually asserts.

### the_new_category_column_creates_a_category_and_its_column

improvise-2zq step 2: the `+cat` column creates a category and its column, without
leaving the grid.

Driven through real keys. The assertion that earns its place is the last one: the
new column is **usable immediately**, which is the whole point —
`records_columns` filters on `kind.is_regular()` and never requires items, so a
category with no items is a column the moment it exists.

### enter_on_plus_cat_in_the_last_row_does_not_reopen_category_creation

improvise-2zq, found by Copilot on PR #27: creating a category from `+cat` shifted
the grid under the advance that followed it.

`add_category` inserts the new coordinate column **immediately before `+cat`**, so
every column index from `+cat` rightwards moves by one. `CommitAndAdvance` then
computed its advance from `ctx`, whose layout is the pre-commit one (gotcha 23:
`apply_effects` rebuilds only after the whole batch). On the last row `Enter`
wraps to the next column — and the index it wrapped to was `+cat` *again* after
the rebuild.

So the failure is not "lands one column off". It **reopens category creation**: a
user naming two categories in a row never reaches the column they just made, and
gets an unbounded `+cat` loop instead. The last row is the arm that wraps, so it
is the arm tested.

### tab_on_plus_cat_does_not_reopen_category_creation

The same stale index reached `Tab`, which advances right rather than wrapping.
Both directions read the pre-commit layout, so both are covered — one arm of a
two-arm rule is the shape
`doctrine(one_arm_of_a_symmetric_rule_is_evidence_about_that_arm_only)` names.

### the_new_category_column_creates_the_column_but_writes_no_coordinate

Committing the `+cat` name creates the **column** and does not write the
coordinate on the row the cursor was on.

The deliberate answer to the ticket's third question, and the reason is gotcha 1:
setting a coordinate moves *every* measure cell in the row. That is a data
migration across N cells, and attaching it to the same `Enter` that made a schema
change — in an app with no undo — is a bigger single keystroke than the measure
`+` asks for. The follow-up edit is one column away.

### the_new_category_column_refuses_past_the_limit_where_the_user_can_see_it

`MAX_CATEGORIES` is a reachable refusal on this path, and the refusal has to be
**visible**.

`draw_bottom_bar` renders the minibuffer instead of the status line, so a message
written while an editor is open is never drawn. The commit path therefore drops
out of editing *before* running the add, and `AddCategory` aborts the batch on
refusal so the trailing `EnterEditAtCursor` cannot re-open the editor over the
message.

The limit itself is checked in exactly one place — `Model::add_category` — and
nothing here restates it; the test fills the document to the limit by asking for
categories until it refuses.

### naming_a_category_in_records_mode_does_not_eject_the_user

improvise-2zq step 1: naming a category in records mode used to eject the user
from the grid they were standing in.

Records mode is `_Index` on Row and `_Dim` on Column, with every regular category
on `Axis::None`. `View::on_category_added` seated a new regular category on the
first axis no *regular* category held — Row — and **bumped the virtual out of the
way** to get there. So `N`, a name, `Enter` moved `_Index` off Row, records
detection stopped matching, and the grid silently became a pivot.

Driven through the real keys rather than through `on_category_added`, because that
is how the bug was found and it is what makes this evidence rather than a
restatement of the fix.

### i_and_o_diverge_on_an_empty_records_grid

improvise-6sq: `i` and `o` want opposite behaviour on an empty records grid, which
is why `edit-or-drill` and `enter-edit-at-cursor` cannot be merged into one
command. Entering records mode puts every regular category on `Axis::None`, so
`EditOrDrill`'s aggregation check is true unless the cursor is on a synthetic
record — and on an empty grid there is none, so `i` reports "cannot drill". `o`
must still open an edit there (see
`add_record_row_in_empty_records_view_creates_first_row`), so routing it through
the same drill check would break it.

improvise-ct5 changed the *message* and deliberately not the divergence. The
outcome — `i` opens no editor where there is nothing to edit — is right, and `o`
one key away is the affordance for adding a row. What was wrong is that the
message said "cannot drill", a pivot-mode concept in a mode whose model is "a
table I type into". Making `i` behave like `o` here (the ticket's option (b))
would collapse exactly the divergence improvise-6sq exists to record, so the
assertion still pins that `i` does not enter editing; only the wording assertion
moved, and it now also names what must *not* appear.

### the_empty_grid_message_is_not_the_records_one_in_pivot_mode

The negative arm of improvise-ct5, and the reason the wording is keyed off the
mode the keymap supplies rather than reworded once.

Without this, "say the records sentence everywhere" passes
`i_and_o_diverge_on_an_empty_records_grid` — and it would promise `o`, which adds
no row in pivot mode, trading one misleading message for another.

The fixture is an empty pivot grid that still counts as aggregated: `Region` on
Row with no items makes `cell_key()` `None`, and `Extra` cycled to `Axis::None` is
the regular category that makes `EditOrDrill` take the drill branch.

### blank_model_bootstraps_its_first_measure_from_the_new_measure_column

A blank model has no categories and no measures, so wide format shows exactly the
two creation columns. Typing a name into the new-measure one creates the measure,
which gives it a real column to type values into — records mode stays
self-sufficient for onboarding (improvise-c5v).

improvise-2zq widened that: a blank grid now offers `+cat` beside `+`, so both
halves of a schema can be bootstrapped without leaving the grid. This test still
pins the measure half; the category half is
`the_new_category_column_creates_a_category_and_its_column`.

Inside it, the cursor lands at column 0, which in a blank grid is now `+cat`
rather than `+` (improvise-2zq) — a record needs coordinates before it needs
measures, so that order is deliberate. `o` leaves the test editing column 0, so
the `Esc` before moving to the new-measure column is load-bearing.

### ctrl_w_prompts_for_a_path_and_a_failed_write_reports_it_where_it_shows

improvise-lae end to end: `Ctrl-W`, a path, `Enter`.

The failing case is the interesting one, and it pins an **ordering** (found by
Copilot on PR #27). `SaveAs` aborts the batch on a failed write, so whatever
`CommitSave` emits after it is skipped. A first version emitted
`[SaveAs, ChangeMode(Normal)]`, which left the app in `SavePrompt` — and
`draw_bottom_bar` renders the minibuffer *instead of* the status line, so the
"Save error:" the effect had just set was never drawn. The save appeared to do
nothing whatsoever.

Leading with the mode change puts the error in `Normal`, where the bottom row is
the status bar. This test therefore asserts the mode *left* the prompt, which is
the opposite of what it asserted before — deliberately, and it is the only thing
keeping the two effects in that order.

### every_minibuffer_mode_entered_from_records_returns_to_records_normal

Copilot round 2 reported that `Ctrl-W` from a records grid returns to `Normal`,
losing the records keymap while the layout is still a records layout. It is right,
and it is a **class**: every minibuffer mode reachable from `RecordsNormal` exited
to `Normal`, because the exit was a constant baked into
`TextEntry::new(..., "normal")` and into each commit command.

The class **predates** `SavePrompt`: `:` and `Ctrl-E` do exactly the same thing
and always have. That is why the fix was not "make `SavePrompt` remember its
origin" — that would leave two of the three arms broken, which is the shape
`doctrine(one_arm_of_a_symmetric_rule_is_evidence_about_that_arm_only)` names.

The binding *table* is not involved and was checked rather than reasoned about: an
`IMPROVISE_BINDING_DUMP` diff shows `RecordsNormal` gains `Ctrl-W` and loses
nothing, so `with_parent` inheritance (gotcha 13) is intact. What was lost is the
runtime mode.

improvise-sihr fixes it by **observation rather than memory**, which is what
took the exit out of the keymap entirely: `App::apply_effects` calls
`settle_normal_mode` after `rebuild_layout`, so the mode a batch *ends* in
matches the grid it ends on and nothing carries an entry mode. The ticket
proposed the opposite — carrying the entry mode on `MinibufferConfig` — and that
was declined for a concrete reason beyond doctrine: carried generically it would
also return `CategoryAdd` to `CategoryPanel`, a behaviour change nobody asked
for, while carried for only three modes it is the one-arm shape again.

This test still drives **both exits** — Esc and Enter — where the original drove
only Esc, because they reach the mode through different effects
(`ChangeMode` and `ReturnToNormalUnlessMoved`) even though one settle now covers
both.

**The first version of the fix resolved per-effect and was wrong in a new
direction**, which the adversarial review of PR #91 caught. Set at the moment
the mode is written, the resolution reads the batch's *opening* layout (gotcha
23) — so a batch that **crosses** settled against the grid it started on. See
`a_batch_that_leaves_records_and_then_sets_normal_lands_in_normal`.

Injected fault: `settle_normal_mode()`'s call site removed. Exactly three tests
fail across the whole workspace — this one, the `:`-line sibling and the
crossing test — which is the per-edit coverage question answered by there being
one edit.

### a_colon_line_run_from_records_settles_back_into_records_normal

The sibling above cannot see `ReturnToNormalUnlessMoved`, and that was found by
injection rather than by reading: forcing *only* that effect back to
`AppMode::Normal` left the sibling green. An **empty** `:` line — which is what
pressing Enter straight after `:` gives — takes `ExecuteCommand`'s
`raw.is_empty()` branch and settles through `ChangeMode`, so the effect that
settles every line that actually runs was untested.

So this test types a line. `page-next` rather than `yank`, because the *script*
parser requires coordinates for `yank` and an unparseable line takes the error
branch — which settles through `ChangeMode` too, and would have made this test
a second copy of its sibling while looking like a new case. The setup assertion
that `parse_line_with` accepts the line is there for exactly that: a future
rename of the command would otherwise turn this test vacuous silently.

Injected fault (against the per-effect version of the fix):
`ReturnToNormalUnlessMoved` forced to `AppMode::Normal`. This test failed,
`left: Normal right: RecordsNormal`, while the sibling stayed green — which is
the point, and is how this test came to exist. Against the shipped end-of-batch
settle the discriminating injection is removing `settle_normal_mode()`'s call
site, which fails this test with the same values.

### a_batch_that_leaves_records_and_then_sets_normal_lands_in_normal

**New wrong state introduced by PR #91's first attempt**, found by the
adversarial review and verified by reverting only `ui/app.rs` and
`ui/effect.rs` to base and re-running the identical probe:

| `:toggle-records-mode . enter-mode normal` | base | #91 round 1 |
|---|---|---|
| from a pivot grid | `Normal` (the sihr bug) | `Normal` still |
| **from a records grid** | `Normal` (correct) | **`RecordsNormal`** |

The cause is gotcha 23: resolving at the moment the mode is *written* reads the
batch's **opening** layout, so a batch that crosses settles against the grid it
started on. `:toggle-records-mode` is the case the round-1 prose offered as
proof that the staleness was a *property* — and it is not, because
`ToggleRecordsMode` emits `change_mode(RecordsNormal)` explicitly on entry and
`ViewBack` restores the pushed mode on exit, so the resolution's answer is
overwritten in both directions and the case exercises nothing. That is a careful
comment standing in for a test, in the PR that recorded
`doctrine(an_injection_that_does_not_fail_names_an_untested_path)`.

The fix moves the resolution to **one** place, `settle_normal_mode` after
`rebuild_layout`, where there is no crossing case left to get wrong.

This test drives **both directions** deliberately. One direction alone is
satisfied by a settle that only ever answers one way, and the pivot→records
direction was *also* wrong in round 1 for the mirror-image reason.

### a_wizard_abandoned_over_a_records_grid_returns_to_records_normal

`CloseWizard` writes `view_state.mode` directly rather than through an effect
that could resolve it, so under round 1's per-effect design it was a **third
writer** needing its own edit — and reverting that edit alone left the whole
workspace green, which the review filed. The end-of-batch settle removes the
per-edit question entirely (there is one edit), and this test covers the path
the review named: the wizard abandoned over a records grid.

It drives `CloseWizard` as an effect rather than typing `:import`, because the
wizard needs a real CSV and the contract under test is one batch's settling, not
the import pipeline.

### a_failed_colon_write_lands_in_normal_where_its_error_is_drawn

A failing `:` command must not strand the user in `CommandMode`.

Found by Copilot on PR #27, and it is the sharper form of gotcha 21. The abort is
*supposed* to stop the chain — `:w /bad . q` must not quit. What it must not also
stop is the **bookkeeping**: the mode restoration that `ExecuteCommand` appends.
Aborting before that ran left the app in `CommandMode` with the stale command line
still drawn, and `draw_bottom_bar` renders the minibuffer *instead of* the status
line — so the "Save error:" set one effect earlier was invisible. A failed save
looked exactly like a key that did nothing.

The answer is **ordering, not classification**: a batch puts its tidy-up where an
abort cannot reach it, which is in front. No effect needs a new "runs anyway"
flag, and the rule generalises to any command composing bookkeeping with something
that can fail.

Both arms are here because they fail in different places: no path at all is
`Save`'s guard (aborting since improvise-8tk, so this arm was broken before that
branch), and a bad path is `SaveAs`'s write (aborting since f0f400d).

### a_failed_colon_write_still_refuses_to_quit_after_it

The other half of the same rule: the abort must still stop the *chain*.
`:w <bad> . q` is improvise-8tk's data loss spelled on one `:` line, and moving the
mode restoration to the front must not have bought the fix above at the price of
that one.

### a_sequence_whose_first_step_aborts_does_not_run_the_second

Gotcha 21 claims a failed `Save` aborts across `Binding::Sequence` steps as well
as within one command's effect list. Nothing tested it: no shipped `Sequence` has
a first step that can abort, and `ZZ` is a *prefix* resolving to the single `wq`
command, so `prefix_z_then_z_reaches_save_and_quit` exercises the
within-a-command case only. The claim was structurally true and unguarded.

It is true because `Keymap::dispatch` flattens a `Sequence` into ONE effect vec and
`App::apply_effects` runs that as one batch, breaking on `abort_effects`. It is
*load-bearing*, because it is the reason a `[w, force-quit]` Sequence would be
safe — the argument the grooming note unblocked on.

And it is exactly what improvise-0hb would remove. Command-level transactionality
means applying effects *between* steps, and `apply_effects` resets
`abort_effects` at entry and again on abort, so step 2 would be dispatched into a
fresh batch with the flag cleared — the `Quit` after a failed save would land,
silently reopening the improvise-8tk P1 data loss. This is the guard that would
catch it.

### unmatched_key_after_prefix_is_swallowed

A prefix is consumed whether or not the next key matches, so a stray key after `g`
does nothing rather than being treated as a fresh normal-mode binding.

### renaming_a_coordinate_moves_every_measure_in_the_row

improvise-c5v: a records row groups one cell per measure, so editing a category
column has to move every one of them. The long-format code moved only the cell
under the cursor, which in wide format would tear the row apart — half its
measures under the old coordinate, half under the new one.

### leaving_records_mode_cleans_empty_key_cells

improvise-3zq (bug #2): `AddRecordRow` creates a cell with an empty `CellKey` when
no Page-axis categories supply coords — that cell serialises as ` = 0` in
`.improv` and re-appears on every records toggle. Leaving records mode must clean
up any such meaningless records (the inverse of the `SortData` that runs on
entry).

### enter_at_bottom_right_of_records_view_exits_editing

improvise-3zq (bug #1): `Enter` on the bottom-right cell of records view should
commit and leave edit mode. Previously `CommitAndAdvance` pushed an
`EnterEditAtCursor` effect unconditionally, so the cursor stayed put and editing
re-opened on the same cell.

#### Helper — `records_model_with_two_rows`

Not a test — the fixture for every Tab/Enter boundary test in this file.

A records-mode app with two data rows, for testing `Tab`/`Enter` behaviour at
boundaries. In wide format a row is one coordinate tuple, so the two rows are two
`Region` items, each carrying both measures (improvise-c5v).

### tab_on_bottom_right_of_records_inserts_below

improvise-hmu: `Tab` on the bottom-right cell of records view inserts a new record
below and moves to the first cell of the new row in edit mode.

#### Note — the `wizard_key_tests` module

Behavioural coverage for the import wizard's key handling (improvise-2wm). The
wizard is driven entirely through `handle_key`, so these tests pin the
user-visible contract independently of whether the step/key decisions live in an
effect or in a command.

### name_model_typing_edits_the_name_and_enter_confirms_the_import

`Enter` at `NameModel` builds and parks on the `Done` confirmation; a second
`Enter` there installs the workbook and closes the wizard (improvise-907, from
improvise-c6j).

### review_proposals_blocks_when_no_category_is_accepted

BUG (improvise-907, from improvise-gfq): an import whose fields all classify as
`Measure` used to walk the user through `DefineFormulas` and `NameModel` before
`build_model` refused. The refusal now happens at `ReviewProposals`, where `c` can
still change a field's kind.

### left_walks_back_through_the_steps_taken

BUG (improvise-907, from improvise-o1c): reaching a later step and wanting to
change a field kind meant `Esc`, which discarded every decision. `Left` walks back
through the path actually taken (gotcha 22).

### done_esc_returns_to_name_model_and_drops_the_built_model

`Esc` at `Done` goes back to `NameModel` rather than discarding the import, and
the workbook built for confirmation is dropped so a later accept cannot install
stale field decisions.

#### Note — the `basis_tests` module fixture

The two lower tiers of a batch's basis (ADR 001, improvise-xwi6). Fixture note:
two regular categories, two measures and two records, deliberately. A one-category
fixture cannot tell `TransposeAxes` from `CycleAxis`, a one-record fixture cannot
tell "the record it touched" from "every record", and a one-measure fixture cannot
tell a record from a cell — the three distinctions these tests exist to draw.

### a_records_two_measures_share_one_record_key

The record is the coordinate tuple with `_Measure` gone, and both measures of one
coordinate name the same record, which is the whole point of the unit.

### a_layout_change_moves_the_layout_version_and_no_record

Tier 2 moves, tier 3 does not. A shape change invalidates every in-flight
`CellData` batch through `layout_version` alone; it must not also masquerade as a
write to any record.

Fixture repaired under improvise-75ld: it asked for
`SetAxis { category: "Month", axis: Axis::Column }` against a fixture where
`on_category_added` had already seated `Month` on Column — so the "layout change"
was a no-op, and the test passed only because the tier was bumped off the
declaration. It now asserts the axis before and after, and asks for `Axis::Row`.
This is the failure that turned "reclassify three effects" into "stop reading the
tier off the classification". Injected fault: the same as
`both_shape_counters_move_for_every_reshaping_command` — fails naming `SetAxis`.

### a_loaded_document_is_dirty_because_its_path_still_names_the_old_one

**A loaded document is dirty, and that is a behaviour change with a reason**
(improvise-gv9r, raised by the independent review of PR #26).

`LoadModel` is `Model(WholeWorkbook)`, so `apply_effects` raises the flag — where
before that branch `:e other.improv` left the document reporting clean. Clean was
the wrong answer: `LoadModel` does **not** move `model_state.file_path`, so after
it the in-memory document is a different one from the file the path names, and the
bytes on disk and the bytes in memory really have parted company. That is all
`dirty` claims.

It is worth a test rather than a comment because of what it gates:
`autosave_if_needed` and the `:q` refusal both read this flag, so a user who loads
and quits is now asked rather than silently dropping a document whose path is
stale. `:q!` still discards. The path-staleness itself is pre-existing and is
reported separately; this pins only that the flag follows from it.

### a_view_only_batch_leaves_the_document_clean

The negative arm of improvise-gv9r's raise, and it must be aimed at
`apply_effects` rather than at a `Command`: a view-slice command is folded
straight into `ViewState` and never reaches this loop, so a negative written
against the session door would pass however unconditional the raise was. It is a
`ViewOnly` effect **run through the loop** that discriminates.

What it pins: a client's cursor, status line or panel does not put the document
out of step with the bytes on disk, so it must not arm an unsaved-changes prompt
or an autosave.

### a_cell_edit_moves_only_the_record_it_touched

Tier 3 moves for the record written and for no other, and tier 2 stays put: a cell
edit does not invalidate batches that merely share the document.

### a_coordinate_rename_is_one_write_to_each_record_not_one_per_measure

The unit is the record, not the cell. Renaming a coordinate in wide records mode
moves *every* measure cell in the row (gotcha 1), and that is ONE write to the
record it left and ONE to the record it arrived at — not one per measure.

### an_aborted_batch_bumps_nothing_it_did_not_run

Gotcha 21: a batch that aborts halfway must bump nothing after the abort — in
either tier. `apply_effects` accumulates the `Dirty` inside the loop for exactly
this reason, and the versions are read off the same place.

### a_write_that_changes_nothing_moves_no_version

A write that lands the value already there invalidates nobody's basis, so it moves
no version. This is a deliberate choice of semantics — observed change, not "a
`CellData` effect ran" — and the safe direction, because the alternative rejects
batches for nothing.

### the_version_tiers_are_not_in_the_view_state

The counters are shared document state on `App`, never on the portable session
slice (gotcha 30): a client inheriting another's idea of how far the document has
moved is exactly the stale basis this machinery exists to catch.

### a_page_flip_under_prune_empty_invalidates_the_cursor

improvise-7kvw, the case improvise-9mo9 decided the other nine and deliberately left
open. `SetPageSelection` changes which records are visible and not which categories
frame them, so a grid position normally keeps naming the same cell. Under
`View::prune_empty` that is false: which rows survive pruning is a function of the
page selection, so a page flip renumbers the grid while neither `reset_cursor` nor
`cursor_epoch` moved — gotcha 24's failure in both directions at once, the host's own
cursor stale and no client told.

9mo9 did not patch it because `[`/`]` paging is constant navigation and sending the
cursor home on every flip is worse than the hole when `prune_empty` is off, which is
the default. So the reset is conditional on `prune_empty`, and on the selection
actually moving.

The fixture has three regions and two quarters whose non-empty rows differ, so
position 0 names `North` on Q1 and `South` on Q2. The test asserts the row labels
really differ before asserting the epoch moved; the negative arm asserts they really
*do not* differ with pruning off, so neither case can pass vacuously.

Injected fault: drop the reset. The prune arm fails. Injected fault: make the reset
unconditional. The no-prune arm fails.

### a_page_flip_without_prune_empty_leaves_the_cursor_alone

The negative arm of the above, and it is load-bearing for the same reason gotcha 24
gives for `SetNumberFormat`: without it, "reset on every page flip" passes the
positive test and costs the user their position on every `[`/`]`.

### selecting_the_page_already_selected_invalidates_no_cursor

The third arm, and the one the user actually hits. `SetPageSelection` resets the
cursor only when the selection **moved**, and that guard shipped untested: deleting
`moved &&` left the whole workspace green at 1416, which review caught.

It is not an edge case. `PageNext`/`PagePrev` are odometer-style and emit a
`SetPageSelection` for **every** page category on every keypress, so with more than
one page category the non-carrying categories are always no-ops, and a single
one-item category re-selects the same item every time. Unguarded, each such keystroke
bumps `cursor_epoch`, sends the host's cursor to the origin and broadcasts
`Command::ResetCursor` to every client on that view — improvise-75ld's "a no-op that
refused everybody else", one tier down and aimed at the cursor instead of the basis.

The test asserts both preconditions **before** applying the effect — that the fixture
really is on `Q1`, and that the grid really does prune — and that the row labels did
not move before asserting the epoch stood still. So it cannot pass by re-selecting a
page that was never selected, or by testing a grid that does not prune.

Both `before` placements are load-bearing, and both were wrong once. The page check
used to run **after** the effect, which sets the page to `Q1`, so it could not see a
fixture that started elsewhere: with the fixture moved to `Q2` the test did fail, but
at the *row-labels* assertion, under a message about renumbering. The sentence naming
the property was not the assertion guarding it. The `prune_empty` check did not exist
at all, and with pruning off the reset never fires whatever `moved` says — so
`paged_model(false)` plus the deleted guard left the test **green**, proving nothing,
with nothing saying so.

Both preconditions read `App::view()`, the view the effect writes through
`app.view_mut()`, and not `Workbook::active_view()`, which is the document's
*default* and is deliberately allowed to differ from it (gotcha 29). They agree in
this single-client fixture, so the distinction is latent rather than a live bug —
but an assertion about what an effect did must read the view that effect wrote.

Injected fault: delete `moved &&`. Fails by name, alone.
Injected fault: `paged_model(false)` **and** delete `moved &&`. Fails on the
`prune_empty` assertion; before that assertion existed, this combination passed.
Injected fault: `paged_model`'s fixture starts on `Q2`. Fails on the page-selection
assertion, which now names it; before the move it failed on the row labels instead.

### a_foreign_layout_is_not_shown_the_hosts_staged_drill_text

improvise-pubt, filed with an argument for unreachability that the test disproves.
`project::cell_display` fills `edit_text` from `App::display_value_in(layout, key)`,
and since improvise-jytm that layout may be another client's.
`drill_state.pending_edits` are keyed by `(row index, RecordCol)` against the
**host's** drill layout (gotcha 1), and a row index in one layout names nothing in
another.

The ticket read the arm as unreachable because "a foreign layout built by
`GridLayout::new` carries no frozen records, so it should not produce a synthetic
key". That is wrong: records mode **auto-detects** from the axes (gotcha 1), so an
ordinary saved view with `_Index` on Row and `_Dim` on Column gives `GridLayout::new`
a records layout whose `cell_key` returns exactly the synthetic `(_Index, _Dim)` key
the drill arm matches on. The test fails on the original code, naming the host's
staged text read out of a layout it was never staged against.

The fix is `App::staged_drill_edit`, which consults `drill_state` only when
`std::ptr::eq(layout, &self.layout)`. A pointer comparison rather than a name
comparison because `grid_in` reuses `&app.layout` by reference where the view names
agree and derives a fresh one otherwise, so identity *is* the question being asked.
Whether the host's pending drill text should reach any other client at all is
improvise-tom1 and is untouched here.

The positive arm — the host's own layout still shows `STAGED` — is in the same test,
or "never consult the drill" passes, which would blank the TUI's own drill editor.
The drill fixture needs a real `RecordRow`, not an empty `Vec`: the host's layout
iterates the frozen snapshot, so an empty one gives a zero-row grid and the positive
arm cannot be built.

### a_default_that_names_nothing_says_which_invariant_broke

improvise-h074. `App::resolve_view` falls back to `workbook.active_view` for a name
the document has not got and then indexes `workbook.views[default]`. Several of these
call sites previously went through `Workbook::active_view()`, which carries
`.expect("active_view always names an existing view")`; `IndexMap`'s `Index` impl
panics with no message at all, so a workbook whose default names nothing — reachable
only by constructing the struct directly, since `create_view`/`delete_view` maintain
it — failed with a bare index panic where it used to name the broken invariant.

Injected fault: restore `&workbook.views[default]`. The test fails on the expected
substring.


---

## `ui/project.rs`

### a_window_reaching_the_last_row_carries_the_trailing_group_header

A collapsed group at the bottom of the grid has no items to hang off, and is still
the only row that says those rows exist. A window that reaches the last row
therefore carries it.

It must be *seen*; it is not yet actionable in a browser. `web/app.js` registers no
listener on a group `<th>`, and `toggle-group-under-cursor` is still in
`the_command_names_that_still_need_the_host_are_these`, so `z` resolves `Upstream`
into `Host::handle`'s empty vec — a black hole, not a slow path
(`doctrine(an_upstream_key_is_a_black_hole)`). The TUI can expand it today; the
browser can only show it.

#### Helper — `app_with_a_trailing_group`

Not a test — the fixture for that test and for its stops-short arm.

`Region` on Row with `East` in a group that comes **last**, so the group header is
the final entry on the axis and nothing follows it. That is the shape
`axis_frame`'s trailing rule is about.

### a_window_that_stops_short_does_not_carry_what_is_below_it

The other arm, and the one that would make a browser scrolled to the middle of a
document draw the bottom of it: a window that stops short of the last row carries
nothing below its last item.

### a_window_opening_inside_a_group_names_only_the_nearest

A window opening *inside* a group still names it — otherwise a browser could not
say which group a column or row belongs to — while `body_from` declines to paint a
header that sits above the window.

**And only the nearest one comes.** The fixture puts **two** headers above the
window (`Hidden` is collapsed, so it emits a header and no items, and `Late`'s
follows it immediately) because that is the only way to tell "the nearest header
comes" from "every header above comes". With one group above, deleting the
compaction leaves this green — it is a *cost* property, O(the groups a deep scroll
crossed), and correctness alone cannot see it: `group_of` still finds the nearest
and `body_from` still starts at the first item either way.

#### Helper — `app_with_consecutive_headers`

Not a test — the fixture for that test and for `consecutive_headers_between_two_rows_all_come`.

Two consecutive group headers between two rows: `Hidden` is collapsed, so it emits
a header and no items, and `Late`'s header follows it immediately.

### consecutive_headers_between_two_rows_all_come

Headers between two in-window rows all come, not just the nearest.

Only the nearest wins `group_of`, so a compaction that kept one would still answer
every question about *membership* correctly — and would silently lose the `Hidden`
header, which is the only thing on screen that says those rows exist at all, and
the only thing that *will* be clickable once the browser can toggle a group. It
cannot today: `z` is `Upstream` and answered with nothing.

Above the window the rule is the opposite, and that arm is
`a_window_opening_inside_a_group_names_it_without_painting_it`.

### folding_what_the_server_emits_reproduces_what_the_server_has

The contract of the whole module: a client that folds what the server sends ends up
holding what the server has. Not the fold axiom — the yardstick (`server_truth`) is
built independently of the projections.

#### Helper — `server_truth`

Not a test — the yardstick that test and its siblings measure a fold against.

What the server believes the client should be holding, built straight from the
authoritative state — the yardstick every fold is measured against, and built
independently of the projections so the fold tests are not vacuous
(`doctrine(fold_tests_must_test_order_not_the_axiom)`).

### removing_a_category_empties_the_clients_formula_panel

Removing a category removes every formula that targets it —
`Model::remove_category` ends with `formulas.retain(|f| f.target_category != name)`.
`RemoveCategory` is `LayoutGeometry`, and `Dirty::of(LayoutGeometry)` reported
`formulas: false`, so the formula panel was never told: a client's `FormulaCache`
kept showing a formula the document no longer has, indefinitely, with nothing to
bring it back short of a reconnect. (Gotcha 35.)

### a_late_client_converges_with_one_that_followed_from_the_start

Bootstrap is a compaction, not a replay: a client that connects late and folds only
the seed lands exactly where one that followed the whole stream is.

### the_projected_extent_is_the_whole_grid_not_the_window

The extent is the **whole** grid, not the window (improvise-bpq2).

A client resolving its own cursor keys clamps against this, so a producer that read
it off the projected ranges would stop every browser cursor at the bottom of the
screen — and would look right in the TUI, whose cache covers the whole grid at
offsets zero (gotcha 44). The fixture is deliberately **not square** (3 rows, 2
columns) and the window is smaller in both directions, so it separates "the whole
grid" from "the window" *and* from "rows and cols swapped".

### a_reshaped_grid_folds_to_the_desired_state

Reshaping is the case the ordering exists for. Growing is the half that bites:
cells at positions the client does not have yet are dropped, so the labels that
create those positions must go first. Shrinking is here for the other direction —
cells that no longer fit must not be sent into a grid that has already been cut
down.

### a_renamed_column_whose_total_did_not_change_still_gets_its_total_back

improvise-hkt6, found by Copilot on PR 20. The diff must carry the mirror to
`desired`, and it stopped doing so for one shape: a column **renamed** while its
total stayed the same number.

`apply` blanks the total under a relabelled column (gotcha 33(b)), so the client's
total went blank; but the diff compared `mirror.totals` against `desired.totals`,
found them equal, and sent no `Totals`. `ClientSession::project_diff` then assigns
`self.grid = desired`, so the server believed the number was still on screen — a
cache wrong **forever**, with nothing in the stream to correct it, which is
precisely the direction gotcha 35 keeps `Dirty::of` sloppy to avoid.

The assertion is the one `a_reshaped_grid_folds_to_the_desired_state` makes — fold
the updates onto the mirror and require the result to be `desired` — because "a
`Totals` update was emitted" would pass for a producer that emitted the wrong
totals.

The fixture renames the **middle** column and keeps every total equal, so it
separates "re-sends what the relabel blanked" from "re-sends whenever the labels
move at all".

### renaming_an_axis_category_resends_every_cell

The producer half of the category-rename rule, and the reason the `apply` half is
not cosmetic: the two compose into "wrong forever" without it.

`grid_updates` diffs the cells against the **framed** mirror — the mirror with the
frame updates already folded in. So an `apply` that kept the cells through a `cats`
change made the producer agree they matched and send nothing,
`ClientSession::project_diff` then assigned `self.grid = desired`, and the client
held cells answering the old coordinates with no update in the stream that would
ever correct them.

An axis swap is the case that bites: two categories can legally share an item name,
so `Region = [North]` becoming `Department = [North]` leaves the labels identical
and the `CellId`s **valid and pointing at the other record**.

### a_session_that_switches_views_is_projected_the_view_it_switched_to

The original bug: `ClientSession::apply` folded straight into the session's mirror,
so a `Command::SwitchView` moved the mirrored `active_view` while the workbook's
stayed put (gotcha 29). `viewport()` then read the cursor under the new view's key
and the next `project()` built the grid from the *other* view.

improvise-jytm answers it the other way round — the mirror is now authoritative and
`grid_in` is built from it — so the assertions moved with it: the asker's mirror
names "Second", the document's default is still "Default", and the rows it is sent
are the ones "Second" has. The fixture's two views are transposes of each other, so
"was I sent the right grid" cannot be answered by accident.

Two faults injected 2026-09-20, both of which fail it. Pointing `grid_in` at
`app.view_name()` regardless of the session reproduces the original defect exactly.
Removing `ClientSession`'s `projected_view` observation is the *new* way to
reproduce it: a switch is `ViewOnly`, so nothing in `Dirty` carries it and the
session must notice its own move or never rebuild its grid at all.

#### Helper — `app_with_two_views`

Not a test — the fixture for every lockstep and rebase test in this file.

A second view whose axes are the first one's transposed, so the two grids cannot be
confused for each other: "Default" rows are regions, "Second" rows are measures.

### a_view_switch_moves_only_the_switchers_mirror

The successor to `a_view_switch_moves_every_sessions_mirror_not_just_the_switchers`,
which pinned the lockstep improvise-jytm removes. The model is shared and the UI is
not, so a bystander keeps its view, its scroll and its grid.

Three assertions, because the old bug can come back through three different doors:
the bystander's mirrored `active_view`, its `row_offset` (nonzero, so a mirror reset
shows as a move rather than coinciding with the origin), and whether `project_all`
names it at all. That third one is the sharpest — a bystander repainted for
somebody else's switch is invisible in the first two, since it would be repainted
with its *own* view's content and look right.

The scroll is drained through `project_all` before the switch, so the fan-out the
test reads names only what the switch caused.

Fault injected 2026-09-20: restoring the `adopt_active_view` loop (push the accepted
view into every mirror) fails it, as does removing the `projected_view` observation.

### a_refused_switch_corrects_the_asker_and_nobody_else

improvise-6c6q. `apply_switch` relayed `Broadcast::ToOthers`, so on the refused path
every client *but* the one that was wrong was told the accepted name. The asker's
own slice kept the name it had folded optimistically while its server-side mirror
held the accepted one — gotcha 24's mismatch, arriving through the asker's own
keystroke. `apply_switch`'s doc comment claimed the hole was closed, which is worse
than the hole.

**The test drives the asker's own `ViewState`**, folding the command the way a real
client would before handing it to the door. That is the whole cost of the ticket:
the sibling bug in `apply_model` shipped green because its test read a
"what do the others hear" helper, which looks away from the asker by construction.
There is deliberately no such helper in this test.

**The bystander is asserted from a view it can be moved OFF.** A first version left it
on the default, where a wrongly-addressed relay folds to a no-op — so the assertion
read as coverage and was vacuous, which review demonstrated by injecting `ToAll` and
watching this test pass. It now sits on "Second", and the audience itself is asserted
(`Broadcast::To(asker, …)`) rather than inferred from what somebody's slice folded
to: the audience **is** the message (gotcha 36), so it is the thing to read.

Two faults injected 2026-09-20, both of which fail it: relaying `ToOthers` (the
original improvise-6c6q bug) and relaying `ToAll` (the over-correction). The second
also fails `folding_the_relay_puts_a_refused_asker_back`, which reads frames rather
than folded state.

### deleting_the_view_this_client_is_on_moves_it_to_the_default

The host-side half of `deleting_a_view_moves_the_client_that_was_on_it_and_tells_it`,
and a desync that predates improvise-jytm: `Workbook::delete_view` moved the
document's active view and nothing moved `ViewState::active_view`, so the TUI's
cursor was read under a key naming a view that was gone. It was invisible while
`draw_status` read the workbook's copy; improvise-k69y makes it read the client's,
which is what turns it into a wrong badge.

The negative arm is the discriminating one: deleting a *different* view must leave
this client where it is. Without it, unconditionally seating the client on the
default would pass the positive assertion. The fixture therefore carries three
views, and the first delete is of the one nobody is on.

Fault injected 2026-09-20: dropping the `view_state.active_view` line from
`effect::DeleteView` fails it.

### deleting_a_view_moves_the_client_that_was_on_it_and_tells_it

A hole improvise-jytm opens and must therefore close. Before the split,
`Workbook::delete_view` moved the document's active view and `adopt_active_view`
pushed the result into every mirror, so deleting the view a client was on moved that
client. The split removes both, and without `adopt_deleted_views` the client is left
naming a view the document has not got: `App::resolve_view` falls back to the
default, so it is *projected* the default's grid while its slice says otherwise, and
`viewport()` reads its cursor under a key nothing will ever write again.

The fixture carries **three** views and the bystander sits on the one the delete does
not remove. That is the whole discrimination: a first version left the bystander on
the default, where a relay addressed to everybody folds to a no-op, and review
demonstrated the vacuity by replacing `Broadcast::To` with `ToAll` and watching the
test pass. The evidence entry had even *stated* the risk in one sentence and claimed
the assertion covered it in the next — the exact shape this file exists to stop.

So four assertions, over three different surfaces. The **audience** is read off the
`Broadcast` itself, because the audience is the message (gotcha 36). The client that
lost its view moved, in its mirror *and* in its own slice, the latter driven by
folding what the host relayed. And the bystander is still on "Third" — in its slice
and in its mirror — which only means something because there is somewhere for a
wrong relay to have moved it from.

The deletion itself is asserted, because `Workbook::delete_view` refuses to remove
the last view: a fixture with one view would pass every other line here while
deleting nothing.

Two faults injected 2026-09-20, both of which fail it: dropping the
`adopt_deleted_views` call from `apply_model`, and addressing its relay `ToAll`.

### a_session_on_another_view_is_projected_that_views_axes

The hole review found in improvise-jytm's first cut: only the **grid** half of
"every projection follows the session's view" was checked. `category_tree_in` and
`tile_bar_in` are reached through `ClientSession::view_in`, and replacing that
function's body with `app.view()` — the host's view — left the whole suite green.

The harm is worse than a wrong pixel. A category's axis is what the **tile bar** is
for: it is the control a user reassigns axes with. Under the fault a client on
another view sees the wrong axis for every category, in the widget it would use to
change them, while the grid beside it is correct — nothing to notice and nothing
that fails.

`app_with_two_views` is what makes two assertions enough: "Second" is "Default"
transposed, so reading the wrong view **swaps** Row and Column rather than blurring
them, and the test names both categories rather than one.

Fault injected 2026-09-20: `ClientSession::view_in` returning `app.view()` fails it.

### walking_the_view_history_moves_no_document_state

The other hole review found: of the three effects improvise-jytm reclassified to
`ViewOnly`, only `SwitchView` had a guard (`a_wire_view_switch_leaves_the_document_clean`).
Putting `ViewBack` or `ViewForward` back to `Model(LayoutGeometry)` left all 1313
tests green — which is gotcha 3's whole point: a mis-declared kind is silent.

It asserts the **consequences** rather than the declaration, because those are what a
reader can price. `App::apply_effects` raises `model_state.dirty` for any
`Model(_)`, and the browser worker persists on that flag — so a mis-declared
view-back rewrites the entire OPFS document (gotcha 41: `Storage` is whole-document)
for a move that changed nothing. And it bumps `layout_version`, so every *other*
client's next `Command::Model` is refused `LayoutMoved` (ADR 001) because somebody
pressed a navigation key.

The `assert_ne!` on the active view before each pair of assertions is the
check-the-checker: an empty history stack makes `ViewBack` a no-op, and a no-op
dirties nothing whatever its kind.

Both effects are driven, because they are separate `kind()` implementations and
review injected them separately.

### a_fresh_session_opens_the_documents_default_view

The other half of improvise-jytm's contract, and the reason `Workbook::active_view`
survives at all: it is what a freshly connected client opens. The fixture switches
an *existing* session first, so a `connect` that read any client's view rather than
the document's would be visible.

Fault injected 2026-09-20: pushing the accepted view into every mirror in
`apply_switch` moves the document's default nowhere, but making `apply_switch` write
`workbook.active_view` fails it.

#### Helper — `diff`

Not a test — the accessor every diff-subject test in this file reads its answer through.

The two halves of an `Applied::Diff`, for the tests whose subject is not the
rebase. It panics rather than defaulting, so a command that starts replacing the
document is a loud failure here and not a silent `Dirty::NOTHING`.

*Note:* before the comment-cap wave this helper's doc comment was preceded by an
orphaned paragraph ("One connected client, as its host really drives it…")
describing a type that no longer exists at that site. Dropped.

*Both `a_model_command_that_moves_the_active_view_binds_every_mirror` and
`a_model_commands_view_switch_reaches_the_asker_too` were retired by
improvise-jytm, along with the `to_others` helper whose blind spot the second one
existed to cover. No `ModelCommand` moves anybody's active view any more:
`ViewBack`/`ViewForward` left the wire because the stack they pop is one client's
history, and a document swap lands in the rebase arm. The property the second test
defended — a correction must reach the client that folded it — is now
`a_refused_switch_corrects_the_asker_and_nobody_else`, which drives the asker's own
slice rather than a helper.*

### a_geometry_command_invalidates_every_clients_cursor_including_the_askers

`TransposeAxes` and `CycleAxis` call `app.reset_cursor()` because renumbering every
row and column makes the old indices name different cells (gotcha 24) — but that
reset only reached the authoritative `App`'s own view slice. Each `ClientSession`
keeps its own cursor in its mirrored `ViewState`, and nothing reset those, so every
client was left indexing a grid that had changed shape under it: `Viewport::window`
clips, so what a bystander actually saw was a **blank** grid, until it happened to
scroll (improvise-gvnb).

The asker is checked alongside the bystander, and that is not symmetry for its own
sake. A `Command::Model` is the one kind a client cannot fold optimistically, so
the asker is in exactly the same position as everyone else, which is why the reset
is `Broadcast::ToAll` and not the `To(id, …)` a refused `SwitchView` correction is.

### a_model_command_that_does_not_reshape_the_grid_leaves_cursors_alone

The other arm. A model command that leaves the grid's shape alone must not throw
anyone's cursor away: where a client is looking is its own UI, and only a change to
the shared geometry may invalidate it. Without this, "reset on every model command"
would pass the test above.

### a_bystanders_own_slice_follows_the_invalidation

And the bystander really is shown the new grid from the origin, rather than a
window clipped to nothing. This is the half the mirror reset alone cannot give: the
client's own slice is driven by what the host relays, so the test drives it the way
`Host::command` does.

#### Helper — `host_dispatch`

Not a test — the relay stand-in that test and its siblings drive a client's own slice with.

A stand-in for `Host::command`'s relay, driving one client's own slice. `to` is the
client being driven and `from` the one that asked, because that pair is exactly
what `Broadcast`'s audience is about.

### a_model_command_that_leaves_the_active_view_alone_broadcasts_no_switch

Renamed in review: it was `..._broadcasts_nothing`, and since improvise-9mo9 it
asserts a `ResetCursor` broadcast, so the name read as a lie about its own
assertion. No client folded an active view, so there is no `SwitchView`
correction to make.
What the others are told is the `Dirty` it produced — and, since improvise-9mo9,
the `ResetCursor` an `AddCategory` owes, because a new category lands on an axis
in every view. The assertion is the exact broadcast list rather than "contains no
`SwitchView`": a weaker form would go on passing if the cursor invalidation were
dropped again.

### every_reshaping_model_command_invalidates_every_cursor

Only `TransposeAxes` and `CycleAxis` reset the cursor, while `SetAxis`,
`HideItem`, `ShowItem`, `ToggleGroup`, `TogglePruneEmpty`, `AddCategory`,
`RemoveCategory`, `AddItem`, `AddItemInGroup` and `RemoveItem` renumbered the
grid and did not (improvise-9mo9). `SetAxis` is the case that shows the shape of
the bug: `CycleAxis` *is* `SetAxis` to `Axis::next()`, so a pair differing only
by an argument disagreed. Because `reset_cursor` is also the only writer of
`cursor_epoch`, every one of these left the host's own cursor stale *and*
broadcast nothing, which gotcha 24 records as strictly worse than doing neither —
the diff then believes each client already holds origin-positioned content and
sends nothing, so they stay stale permanently rather than until the next scroll.

Two faults were injected. Deleting `app.reset_cursor()` from `SetAxis` fails the
bystander-cursor assertion (`left: 1, right: 0`). Changing the fixture so the
item `ShowItem` reveals was never hidden fails the *first* assertion — "ShowItem
reshaped nothing, so this case is not evidence of anything" — which is the one
that stops this from becoming a table of commands asserted against a grid that
did not move. That guard matters most for `AddItem`: the fixture puts two
categories on Row precisely so an append to the inner one inserts a row inside
every outer block, because on a single-level axis an append lands at the end and
moves nothing, and a one-level fixture would have made the reset look unearned.

#### Helper — `app_with_a_reshapeable_grid`

Not a test. Two categories on Row, one of them grouped, one item hidden and one
item left with no data — so `ShowItem` has something to reveal, `ToggleGroup` has
a group to fold and `TogglePruneEmpty` has an empty row to prune. Every command in
the table has something real to reshape, which is what the first assertion checks.

### drilling_resets_only_the_drill_views_cursor

improvise-9mo9 made `SetAxis` call `app.reset_cursor()`, and `DrillIntoCell`
issues eight or so `SetAxis` effects to build the `_Drill` view. The PR claimed
that was safe and **did not test it** — the claim was that `SwitchView` precedes
the batch, so `reset_cursor` writes the *active* view's cursor, which by then is
the freshly created `_Drill` one, while the view drilled *from* keeps its entry
in `ViewState::cursors`. That is an ordering argument about effects inside a
batch, which is exactly the class of claim gotcha 23 exists because people get
wrong.

Two faults injected. Removing the `SwitchView` from `DrillIntoCell`'s batch fails
the **first** assertion — "the drill did not switch views, so this proves nothing
about per-view cursors" — which is the guard that stops the test passing for the
wrong reason if the drill ever stops switching. Changing `reset_cursor` to clear
the whole `cursors` map rather than the active entry fails the third:
`left: (0, 0), right: (1, 1)`, the cursor in the origin view thrown away.

It asserts both sides on purpose. That the `_Drill` view opens at its origin is
half the property; that the *other* view keeps `(1, 1)` and `row_offset: 1` is
the half that would have broken, and a test asserting only the first would pass
against a map that had been emptied.

### an_accepted_no_op_invalidates_no_cursor

The third negative arm, found in review. Gating the reset on `Ok` is not the same
as gating it on "something changed", and six commands the document *accepts*
changed nothing and reset every cursor anyway: `SetAxis` to the axis already
held, `HideItem` on an already-hidden item, `ShowItem` on a never-hidden one,
`AddItem` of an existing item (`Category::add_item` dedupes and returns `Ok`),
and `RemoveItem`/`RemoveCategory` of something absent.

The first assertion is the non-vacuity guard and it is doing real work here: it
uses `grid_shape` to prove each case really is a no-op, so a fixture drifting
into one where the command *does* reshape fails rather than passes.

**Review then found five more cases the table could not reach by construction**,
because `app_with_a_reshapeable_grid` had no formula, no category on `Axis::None`
and no unknown-name case — the one-category-fixture shape the doctrine names. The
fixture now carries a `Profit` formula, a `Channel` category parked on
`Axis::None`, and a `Ghost` entry hidden under a name the model never had, and
the table covers: `SetAxis` on an unregistered category, `HideItem`/`ShowItem` on
a name the grid never drew, `ToggleGroup` on a group that does not exist, and
`AddItem { _Measure, Profit }`.

That last one is the one to remember. The guard it replaced read
`category.items`, but `_Measure`'s effective items are explicit ∪ formula targets
(gotcha 7), so it called a drawn column "new". **Two of the five injections
initially failed to reproduce, and both times the injection was wrong rather than
the guard**: the `ShowItem` case needed a hidden *ghost* name, since
`is_hidden` is false for a name nobody hid; and the `AddItem` case had to compute
its predicate **before** the add, because afterwards the item really is in
`category.items`. An injection that does not reproduce the original ordering
tests nothing — the same shape as the simulated-fix confirmation in gotcha 48.

`ToggleGroup`'s first guard was also wrong and this test caught it:
`add_item_in_group` sets `item.group` without pushing to `Category::groups`, so
asking `groups` said "no such group" about the fixture's real one.

Faults injected, two, both caught by name: removing `SetAxis`'s `moved` check,
and removing `AddItem`'s `fresh` check. The TUI repro needs no test harness —
TileSelect's `r` emits `SetAxis` with no comparison against the current axis.

### naming_a_measure_at_the_bottom_right_leaves_the_cursor_put

A regression this PR introduced and review caught. `AddItem` gained a
`reset_cursor`, and `CommitAndAdvance{Down}` on the trailing `+` column routes
`RecordCol::NewMeasure` through `add_measure` → `effect::AddItem`.
`inserts_a_column` matches only `NewCategory`, so `NewMeasure` falls through to
the advance — and `EnterAdvance` at the bottom-right returns `AbortChain`,
emitting no `SetSelected`. Nothing restored the cursor. The `SetSelected` repair
added for `+cat` covered that arm only.

The fix moves the restore **before** the advance, which is gotcha 21's rule
(tidy-up goes where an abort cannot reach it) applied to the cursor rather than
to the mode. Injecting the old ordering fails it: `left: (0, 0), right: (0, 3)`.

### a_refused_model_command_invalidates_no_cursor

The second negative arm, and the one the first version of improvise-9mo9's fix
got wrong. **It was a loop of one until review**: it named a class and tested
`AddItem` alone, and injecting `reset_cursor()` into `AddCategory`'s `Err` arm —
the very bug it was written for, one effect over — left it green. It is now a
table over all three refusable effects, and that injection fails it by name. `AddItem` and `AddItemInGroup` report a failure — an unknown category
— through `status_msg`, and the reset was written *after* the `if let Err`, so a
refused add bumped `cursor_epoch` and broadcast `Command::ResetCursor` to every
client for a change the document never made. `AddCategory` had been gated on
success in the same commit, which is what made the inconsistency visible.

`SetNumberFormat`'s arm cannot see this: it tests a command that *succeeds* and
reshapes nothing, where this tests one that reshapes nothing because it was
*refused*. Injecting the fault — moving the reset back outside the match —
fails it on the bystander's cursor. The first assertion is the non-vacuity
guard: if the category ever became valid, the case would stop being about a
refusal and the test would say so rather than pass.

#### Helper — `grid_shape`

Not a test. The `row_items`/`col_items` of the layout, as text: which cell each
position names. It is compared before and after rather than counted, because a
`cats` change can re-identify every position without changing how many there are.

### a_clients_view_only_command_moves_only_that_clients_mirror

The other half of the door: a command that is nobody else's business must stay that
way. A client's scroll must not move the authoritative `App`'s own view slice — the
model is shared, the UI is not.

### a_document_swap_reseeds_every_session

Replacing the document must clear the cursor map, because the incoming document's
view names have nothing to do with the outgoing one's (gotcha 24).
`install_workbook` does that for the `App`'s own view slice, but nothing touched the
session mirrors: a connected client went on naming the old document's view and
scrolling by the old document's cursor, so it was projected a window past the end of
a document it had never seen — an empty grid, with nothing that would ever bring it
back.

#### Helper — `other_document`

Not a test — that test's incoming document, and the one whose rename hides a bug.

A different document, with a differently-named active view, so a mirror still keyed
to the outgoing one is visible rather than coincidentally right.

### rebase_hands_every_session_a_snapshot_of_the_new_document

The reseed is reached from the projection path itself, not from a hook a host must
remember to call — `Sessions` is not reachable from the effect that installs a
workbook, so a convention there would be unenforceable. `rebase` exists for the
other half: the client's *own* view slice lives on the client, and only a fresh
`Snapshot` resets it.

### a_document_swap_over_the_wire_rebases_every_client

A document swap asked for by one client must reach every *other* client's own view
slice, and the host must not be able to miss it.

`Command::Model` put the `WholeWorkbook` mutations on the wire (improvise-kmgs), so
`LoadModel` is now something another client can ask for. The server stayed honest —
the generation bump reseeds every mirror on the next `project` — but a bystander
went on holding `cursors["Default"] = row_offset 3` from a document that no longer
exists, and `Applied { dirty, broadcast }` said nothing that would tell its host to
send a `Snapshot`. `Sessions::rebase`, the only thing that produces one, had no
non-test caller in the tree.

That is gotcha 24's cursor-vs-geometry mismatch, arriving through another client's
command — the same shape `doctrine(lockstep_binds_every_session_not_the_asker)`
names, one layer out: the obligation was addressed to a host in prose instead of
being carried by the value it gets back.

The server's own mirror self-heals off the generation bump
(`a_document_swap_reseeds_every_session` covers that half); this is the other one,
the client's slice, which only a `Snapshot` resets.

#### Helper — `a_document_shaped_differently_under_the_same_view_name`

Not a test — the fixture that test needs instead of `other_document`.

A different document whose active view has the *same name* as the outgoing one, and
a different shape: one row where the outgoing document had three.

`other_document` renames the active view, and that rename is what makes the bug
below invisible: `apply_model` broadcasts a `SwitchView` when the name moves, a
client folds it, and the stale cursor is then read under a key nothing has written
yet — right by accident. Same name, different geometry is the case with nothing to
hide behind, and it is the ordinary one: two revisions of a document both call
their first view `Default`. (Gotcha 39's fixture warning.)

#### Helper — `app_over_a_loadable_document`

Not a test — the `App` that makes a wire `LoadModel` in that test really replace the document.

An `App` whose storage already holds that document, so a `Model(LoadModel)`
arriving over the wire really does replace the document rather than setting an
error status.

### a_command_that_leaves_the_document_in_place_is_not_a_rebase

The other arm of the same rule, so the check is not one that always says "rebase"
(`doctrine(one_arm_of_a_symmetric_rule_is_evidence_about_that_arm_only)`). A model
command that mutates the document in place, and a view-only one, are both diffs —
the question `apply_model` asks is whether the generation moved, not which variant
arrived.

### ui::project::tests::a_load_that_fails_is_not_a_rebase

Qualified because `ui::host::tests` has a test of the same name: this is the
`apply_model` one, and a bare leaf was satisfiable by either.

A `LoadModel` naming a document the storage does not hold leaves the document
alone, so it is a diff too. This is why `apply_model` asks the generation rather
than the variant: "is this one of the three `WholeWorkbook` commands" is not the
same question as "was the document replaced", and a client would be reseeded onto
the document it was already holding, losing its cursors to a failed load.

#### Note — the `cell_identity` module

Emitting and resolving the surrogate a client names a cell by (ADR 002). The two
halves are tested together on purpose: an id that is emitted and never resolved, or
resolved by re-deriving it the same way it was emitted, would agree with itself
while agreeing with nothing else.

### the_two_measures_of_a_row_share_a_record_and_differ_by_measure

Gotcha 1's unit, projected: two measure columns of one row are one record and two
cells. Both halves matter — a surrogate that also varied by measure would break
`record_versions`, and one that did not vary by measure at all would make the two
columns the same cell.

### a_projected_record_id_resolves_to_the_cell_it_was_projected_for

The round trip that makes the surrogate usable: what the server projected for a
position resolves back to the key that position has. Every row is checked, because
one row cannot tell "resolves correctly" from "resolves to the first record".

#### Helper — `raw_app`

Not a test — the fixture for every stored-cell surrogate test in `cell_identity`.

`Region` on Row, `_Measure` on Column. Nothing regular sits on None, so every cell
is a stored cell. Three rows and two measure columns, so "the record" cannot be
confused with "the row" or with "the cell".

### a_projected_aggregate_id_resolves_to_the_same_constraint

An aggregate names a coordinate *constraint*, and it must come back as the same
constraint — including the page and filter coordinates baked into the layout's key
(gotcha 27), which is why this resolves against `cell_key` rather than against the
row and column labels.

#### Helper — `aggregated_app`

Not a test — the fixture for the aggregate half of `cell_identity`.

The same, with a second *regular* category left on None, so every visible cell sums
over `Product` rather than being one stored cell. That is the fork `EditOrDrill`
makes, and the one the wire projects.

### a_surrogate_for_a_removed_row_fails_to_resolve_and_the_others_do_not

The first arm of staleness. The identifiers are held constant and only the thing
that makes the id stale is varied (`doctrine(a_rename_can_hide_a_stale_key)`):
"East" is removed, so its surrogate must fail to resolve — while "North"'s, taken
from the same projection, still resolves to the same cell. Without that second
assertion a `resolve_cell` that returned `None` for everything would pass.

### an_empty_cell_can_be_written_through_its_projected_id

improvise-1ey is "type a number into a cell and see the value update", and the cell
a user types into is usually **empty** — so it is named by the grid and by nothing
in the store. The index cannot answer it, which is why resolution falls back to the
layout, and this is the case that makes that fallback load-bearing rather than
defensive.

### a_write_by_surrogate_lands_on_exactly_that_cell

Two rows and two measure columns, so a write that went to "the first cell", "the
whole row" or "the whole column" is visible.

### a_write_through_a_dead_surrogate_changes_nothing

The second half of "reject, nothing to write to": a surrogate that does not resolve
must apply *nothing*, not fall back to some other cell. The document is compared
whole, because a write to the wrong place is exactly the failure being excluded.

It goes through `model_effects` rather than the wire form, because the session door
reaches the same conclusion first and typed — `Checked::against` refuses this
command with `Rejection::RecordGone`, so a `Checked` over it does not exist. What
is asserted here is the **floor underneath** that door: the key path and any host
resolving a command locally get an empty batch rather than a misdirected write.

### a_records_mode_cell_is_named_by_its_coordinates_not_by_its_position

The trap gotcha 1 sets. `GridLayout::cell_key` returns a synthetic `(_Index, _Dim)`
key here, carrying *positions* — the very thing ADR 002 rejected `_Index` for,
because position names a different record after a filter, an insert or a delete.
The projected id must come from `RecordRow::coords` instead, which is what
resolving it back to the real cell proves.

#### Helper — `records_app`

Not a test — the fixture for every records-mode surrogate test.

Wide format: one row per coordinate tuple, a column per raw measure, the coordinate
columns in front and a `+` column behind. Two records and two measures, so neither
"the row" nor "the column" can stand in for the record by accident.

### the_measure_columns_of_a_record_row_share_its_record

Two measure columns of one row are two cells of one record, and the next row is a
different record — the same shape the pivot case has, checked here because records
mode reaches it by a different path.

### a_coordinate_column_is_projected_with_no_identity

A coordinate column is not a cell to write into: editing it renames that coordinate
across *every* measure cell in the row (gotcha 1), which is a different intent from
`SetCell`. So it is projected with no id at all, and a client cannot compose a
write against it.

### a_write_moves_the_version_under_the_clients_own_surrogate

The version counter is keyed by the same surrogate the client holds, which is what
lets one identity answer both ADR 002's question and ADR 001's. A write must move
the counter the client would check.

#### Helper — `apply_caught_up`

Not a test — and note it is `#[cfg(test)] fn`, not `#[test] fn`.

`Sessions::apply` for a client that is fully caught up, for the tests whose subject
is not the basis. The staleness tests build the `Basis` by hand, so neither kind
can pass by accident through the other's convenience.

---

#### Note — the `basis_check` module

ADR 001's basis check: a batch composed against a picture of the document that has
since moved is refused, and one composed against the current picture is not
(improvise-xwi6, improvise-fkfg).

**Both arms, per tier.** A check nobody has watched *reject* is indistinguishable
from one that always accepts, and a check nobody has watched *accept* is
indistinguishable from one that always rejects (`doctrine(check_the_checker)`). So
every tier appears twice.

The fixtures hold every identifier constant and vary only what makes the basis
stale (`doctrine(a_rename_can_hide_a_stale_key)`): the record a batch names is
projected once and reused, so a rejection can never be an artifact of the test
naming something different the second time.

### a_cell_edit_is_rejected_when_the_layout_moved_under_it

**The class-2 case, and the one test to keep if only one survives.**

A `CellData` batch is stale when the *layout* moved, even though no record it names
was written: the coordinates its key hangs off changed underneath it. This is A
removing an item while B commits an edit keyed by it, and it is corruption rather
than a lost keystroke — which is exactly what a cell-level lock would not have
covered, because holding a cell does not stop anyone deleting the row.

#### Helper — `basis_for`

Not a test — the claim every `basis_check` test moves exactly one tier of.

The basis a caught-up client holds for one record. Built by hand rather than
through `caught_up` so each test can move exactly one tier of it and leave the rest
true.

### a_cell_edit_composed_after_the_reshape_is_applied

The other arm of the same tier. A batch composed after the reshape is applied — so
the rule is "the layout moved since you looked", not "the layout has ever moved".

### a_write_is_rejected_when_the_record_it_names_moved

Class 1: somebody wrote the record since this client read it. **Data** staleness —
the record still resolves, so force is coherent and the reason says so. The
identifiers do not move, only the value, so a rejection cannot be a renamed key in
disguise.

### a_write_survives_another_record_moving

The arm that keeps the tier from being "reject whenever anything was written". A
record moving does not invalidate a batch that names a *different* one — which is
the whole reason the tier is per-record rather than a single counter.

### a_write_to_a_record_that_no_longer_exists_says_so

Class 2 by the other route, and the reason ADR 002 put a *maintained* index in
`DataStore` rather than regenerating one from the layout: "the record is gone" and
"the record changed" must not collapse into one answer. There is nothing to write
to here, so nothing to force — force never creates structure.

The test brings the layout claim up to date so that, whatever kind `RemoveItem`
wears, only the record tier is left to answer. It is `CellData` today — removing an
item drops its cells and moves no view's axes — so that line is currently a no-op;
it is kept because a reclassification to `LayoutGeometry` would otherwise reject on
the tier above and hide the case this test is about, silently turning it into a
second copy of `a_cell_edit_is_rejected_when_the_layout_moved_under_it`.

### a_claim_that_omits_the_record_it_writes_is_not_a_free_pass

The force-by-omission hole, closed. Which records are checked is the **server's**
decision, read off the command, so a client that simply left its record out of the
claim does not skip the tier — it is told it never stated one. Without this,
ADR 001's "force is a re-composition, not a bypass" would have a bypass sitting in
the message format.

### a_document_swap_invalidates_every_claim

The top tier invalidates everything, whatever the other two say. It is checked
first so the answer names the largest thing that moved: a client told "a record
moved" after a document swap would go looking for a record that is not there.

### a_rejected_batch_changes_nothing_anywhere

Rejection is a *normal outcome*: nothing is applied, nobody else hears about it,
and no basis moves. A rejection that dirtied something would tell every other
client to redraw for a change that did not happen.

### a_published_basis_follows_the_grid_updates_it_describes

The order `ClientSession::basis_update` claims is load-bearing, pinned.

The basis names record counters for the cells the grid updates in the same batch
carry, so a `Projection::Basis` that overtook them would let a client claim a
version for a value it had not been shown — the mirror image of gotcha 33's rule
that labels precede the cells they frame, and unlike that rule it had nothing
holding it but the order of two statements in `project_diff`. Proved by injecting
the swap (gotcha 43).

### a_client_that_folds_the_stream_composes_an_accepted_batch

The check is only satisfiable because the server *publishes* the basis — which
ADR 001 does not mention, and without which a client could only forge one. So: a
client that folds what it is sent can compose a batch that is accepted, end to end,
with no test-only constructor anywhere in the path.

### re_composing_against_the_republished_basis_is_accepted

And the other side of publication: after somebody else writes, the stream carries a
*new* basis, so the client that folds it can re-compose and be accepted. That is
ADR 001's force — a re-composition against the current basis — reachable without
any mechanism that skips the check.

## `ui/host.rs`

### connecting_hands_that_client_a_snapshot_that_draws_the_document

AC 2: the bootstrap is sent on connect, addressed to the new client, and it is not
empty — a `Snapshot` whose seed carried nothing would satisfy "a snapshot was sent"
while leaving the client blank.

### the_reported_extent_bounds_the_snapshot

The extent is what the snapshot is built at, so a client that says it can see one
row is not sent two. This is the property that makes `Viewport` the handshake rather
than an afterthought.

### without_the_command_the_cell_is_empty

`doctrine(check_the_checker)`: the preceding assertion would also pass if the seed
happened to contain the value already. Inject the fault — never send the command —
and it must fail.

### a_bystander_is_projected_another_clients_edit

A cell edit is the whole document's, so a client that did not ask for it is told
about it too. The projection is per-session because the viewports differ, not
because the change does.

#### Helper — `model`

Not a test — and note it is `#[cfg(test)] fn`, not `#[test] fn`.

A model command from a client that is fully caught up, for the tests whose subject
is not ADR 001's basis. It goes through `handle` like any other frame; only the
basis is supplied rather than folded, and `caught_up` can do that because it holds
the `App` (a client cannot). The staleness tests build their own `Basis`.

#### Helper — `addressed`

Not a test — how every `host.rs` test names a cell the way a client would.

The same cell as a client names it: the opaque surrogate the server projected on
`CellDisplay::id`, which is the only handle a client has (ADR 002). A `CellKey` is
not constructible over there.

### a_scroll_reaches_only_the_client_that_scrolled

A view-only command reaches nobody else. The counterpart to the test above: if every
command fanned out, that one would pass for the wrong reason.

### a_batch_of_commands_is_applied_whole_and_in_order

One keypress resolved on a client is up to three commands, and the host must take
all of them (improvise-bpq2).

The fixture is the shape that actually arrives: a cursor moved past the bottom of a
one-row window, so the selection *and* the offset that follows it travel together.
The assertion is the offset, not the selection, because a host that applied only the
first command would still look right on `selected` — the window is what goes wrong,
and it goes wrong silently.

### a_batch_is_abandoned_at_the_first_rejection

A refused batch applies nothing *after* the refusal. The stronger half — nothing
*before* it either — is `a_refused_batch_leaves_the_commands_before_it_unapplied`;
this is kept because the two halves fail independently.

### a_refused_batch_leaves_the_commands_before_it_unapplied

improvise-955n / improvise-08wb. **The bug**: ADR 001 is emphatic that a stale batch
is never applied, *not even partially*, and improvise-ondj went to the trouble of a
type (`Checked`) to make the unchecked path unconstructible. But `Checked` is **per
command**, and a `ClientFrame::Commands` is a sequence of them: `Host::commands`
applied in order and abandoned at the first rejection, so a frame whose *second*
command was refused had already applied its first. The guarantee was true of a
single command and false of a batch.

The earlier command is a **view** command, which is not a concession: it is the
shape the one real producer emits (`keys::commit` — one mutation among view
commands), and it is still observable, because the host folds it into that client's
mirror. A frame may carry more than one mutation; what it may not do is apply some
of them and refuse the rest, which is clause 4 and open.

**A refusal test needs an accepting one**
(`doctrine(a_helper_that_builds_the_world_the_code_wants_tests_nothing)`: pair a
refusal with an acceptance, or the refusal can pass by refusing everything). The
second half runs the same leading `SetSelected` followed by a command whose basis is
*current*, and requires both to land. Without it, a `SetSelected` that simply never
worked would clear the first assertion and prove nothing at all.

Inside it, the stale claim is built from `caught_up` and then has a *single* tier
bumped, so the refusal can only be about the tier that was moved — a hand-built
basis that is stale in three ways at once cannot tell which check answered.

### a_later_write_in_a_chain_still_answers_for_its_own_record

improvise-955n. **The regression 90ea26e introduced, and the reason clause 2 cannot
be applied to all three tiers.**

Hoisting the basis check to the frame hoisted *all* of it, so every mutation after
the first was admitted with no record check at all. Two clients, one record: A
composes `[SetCell(North), SetCell(South)]` against an honest current basis, B
writes South, A's frame arrives — and A's write to South landed on top of B's
**silently**, which is exactly the lost update ADR 001's record tier exists to
prevent. At the base commit the second command was checked and the frame refused.

**The per-command check was doing two jobs and the commit noticed one.** "Is the
client's picture of the *document* current" is a frame-level fact and hoists
correctly. "Is the client's picture of *this record* current" is not: which records
are at stake is a property of the command, which is why `records_named` is an
exhaustive match over `ModelCommand` (gotcha 43). The incapability argument that
motivated clause 3 is true of records a frame **creates** and was over-generalised
to records it merely **writes** — and those are the ones the tier is for.

Three arms, because a fix that refuses everything would pass the first alone and a
fix that checks nothing would pass the last two.

In arm 2, B moves South with a `SetCell`, which is `CellData`: neither the document
nor the layout tier moves, so only the record tier can discriminate — which is what
makes this a test of the record tier and not of the hoist in general.

**Arm 3 — the same record twice in one chain.** Legitimate and likely (two measures
on one record is one record) and it works for a reason worth pinning rather than
rediscovering: **every check happens before any command is applied**, so the second
write is compared to the record's *pre-frame* counter, not to the one its own
sibling just bumped. Were the checks interleaved with the applies, this frame would
refuse itself with `RecordMoved`.

**Read this before closing clause 4.** Interleaving is tempting there, because it
would close the *silence*: `[RemoveItem(North), SetCell(North)]` under interleaving
is refused `RecordGone` instead of silently half-applying, which is half of what
clause 4 owes. So check-before-apply is not a free win — it buys the legitimate
double-write and **sells** the loud failure on a broken chain, and clause 4 has to
pay that back some other way. Measured rather than argued: injecting the
interleaving fails exactly two tests, this one and
`a_frame_whose_later_mutation_fails_is_not_yet_undone`, and the second failing *is*
the trade. Arm 3 is what stops someone taking the trade without noticing they made
it.

### a_data_dependent_chain_lands_in_order

**The capability the contract is for**: a chain of data-dependent commands, applied
in order, each seeing what the previous one left.

`[AddItem("East"), SetCell(the East record)]` is the archetype, and it is the case a
per-command basis check **cannot** admit however it is ordered: a client cannot
claim a counter for a record its own chain is about to create, because
`BasisCache::claim` filters to records the cache has been told about. So the basis
is answered once, on the first command that carries one, and the rest of the chain
runs against what its predecessors left.

The second assertion is the one that earns the test. Asserting only that the item
exists would pass against a host that applied the first command and dropped the
second — which is exactly what the previous rule did, loudly, and what the pre-frame
check did, silently.

### a_frame_whose_later_mutation_fails_is_undone_whole

**ADR 001 clause 4, delivered (improvise-yqpd). This entry keeps the hole it
characterized, because the inversion is only legible against it.**

The three assertions flipped. The removal is **undone**, the cell reads its
**pre-frame** value — `Some(Number(1.0))`, which is neither `5.0` (the frame's own
write landing) nor `None` (the removal surviving), and that middle value is the
whole difference between a rollback and an abort — and the frame is **refused out
loud** as `Rejection::ChainBroken`. A fourth assertion was added that the previous
version could not have: the **record counter** goes back too, or the client's next
batch is refused for a write that never happened.

**The mechanism, and where it deviates from the ticket.** The ticket asked for
snapshot-and-restore at `App::apply_effects`. What landed is `App::snapshot_document`
/ `App::restore_document` — the workbook, its dirty marker and the two counted tiers,
the record tier riding inside the workbook's `DataStore` — driven from
`Host::commands`, which takes the snapshot before the frame and restores it on
failure. It is **not** taken inside `apply_effects` itself, and the reason is cost:
`apply_effects` runs on every keypress, and an unconditional `Workbook` clone there
would put a full document copy on the held-key path that gotcha 24's `mcq4` work
exists to keep cheap. The primitive is on `App`, so a future `RunScript` opens the
same transaction; only the *trigger* is at the frame.

**What counts as a failure, and the split the ticket did not name.** Two things:
an effect that sets `abort_effects` (`Rejection::MutationRefused`), and a command
whose record no longer resolves (`Rejection::ChainBroken`). The second is the
interesting one. The ticket says interleaving the record check with the applies
would close clause 4's silence for free but would sell the legitimate double-write —
a frame writing one record twice works only because the second write is compared to
the record's *pre-frame* counter. That is true of the **counter** half of the record
tier and **not** of the **structural** half: writing a record twice cannot make it
stop resolving. So the record tier splits — *does it resolve* is asked again before
every mutation (`reduce::broken_chain`), *is its counter what you claimed* stays
pre-frame — and clause 4 gets its loud failure without paying the price the ticket
costed. `a_later_write_in_a_chain_still_answers_for_its_own_record` arm 3 still
passes, which is the guard that the trade was not taken by accident.

**Injected faults, each watched to fail.** Making `undo_frame` restore nothing
fails this test alone. Making `broken_chain` return `None` unconditionally fails
this test and the mirror test.

**A first version of this entry claimed `broken_chain`'s create-exemption was
"correctness-preserving but not currently observable", and that was wrong in both
halves.** Review built the fixture it said did not exist and found the exemption
was *swallowing a write* — see
`a_frame_whose_write_would_land_nowhere_is_refused_rather_than_dropped`. The
exemption is gone from `broken_chain` and stays in the pre-frame check, where it
belongs. The lesson worth keeping: **"no test fails when I remove it" is evidence
the tests are thin, not evidence the code is inert** — the honest next step is to
build the discriminating fixture, and the reasoning that excused not building one
here ("`resolve_record` falls back to the layout, so the case cannot arise") was
an argument about the *common* case presented as an argument about all of them.

---

The hole this replaced, kept because the inversion is only legible against it:

The contract says a frame is all-or-nothing. It was not: `App::apply_effects`
aborts without undoing (gotcha 21 — an abort is not a rollback), so a chain whose
later mutation fails leaves the earlier one standing.
`[RemoveItem("North"), SetCell(the North record)]` is the shape: the removal lands,
the write then resolves to nothing, `model_effects` returns an **empty** batch, and
the frame is reported accepted having applied half of itself. Both the partial
application and the silence are open.

**This test did not pin that, and the way it failed is the repo's own signature
defect sitting on the one test whose job is to fail loudly if anyone believes the
hole is closed.** It asserted that North was removed and that nothing was rejected —
both true of any *wholly successful* frame. The independent reviewer retargeted the
second command from the removed North to the live South, making the frame succeed
completely, and the test still passed. It could not tell half-applied from
fully-applied, so it would never have inverted when clause 4 landed; it would have
gone on passing. The missing assertion is the one that names the hole: **the second
mutation did not land.**

It is unreachable from any client here — the only producer of a multi-command frame
is `keys::commit`, which emits one mutation among view commands — which is what
makes it safe to document rather than rush.

The frame claims North at its current counter, so it passes the check and fails at
*apply* time — which is the only way to reach the hole. A frame refused by the check
never applies anything and would be testing clause 2 instead. That fixture note
survives the inversion unchanged: it is still the only route to the arm.

### a_rolled_back_load_leaves_no_ghost_view_and_no_false_status

**Clause 4 was delivered for the document and the mirrors and NOT for the host's
own session slice**, found by an integration review of the first two commits.
`App::restore` said `view_state` was "deliberately untouched", citing
improvise-yqpd's design note: *`AddCategory` writes its refusal into
`view_state.status_msg` and THEN aborts, so rolling that back would undo the
explanation along with the damage.*

**That note is about an `apply_effects`-level rollback, which is not what
landed.** At the *frame* level the asker receives `ServerFrame::Rejected(reason)`
— the explanation is delivered on the wire, addressed to the client that earned
it — so the host's status line has no explanatory duty for a client's frame. The
note still governs the host's own key path, which aborts without undoing exactly
as gotcha 21 describes, and is untouched here.

Three measured consequences of a refused frame, all now closed: the TUI host's
cursor still moved and `cursor_epoch` was still bumped; a rolled-back `LoadModel`
left `view_state.active_view` naming a view of the **discarded** document; and
`status_msg` kept the rejected frame's sentence, so **a frame that did nothing
claimed it had loaded a file**.

**The fixture had to be renamed before the test could see any of it.**
`host_over_a_loadable_document` built its incoming workbook with
`Workbook::new`, whose default view is `Default` — the same name the outgoing
document uses — so a ghost `active_view` was indistinguishable from a correct
one. This is gotcha 39's trap one level up, and the incoming view is now
`Loaded`. Checked first: neither existing user of that fixture depends on the
names matching, and gotcha 39's own same-name fixture is
`a_document_shaped_differently_under_the_same_view_name` in `project.rs`, which
is untouched.

Why restoring the whole slice is safe rather than clobbering a TUI user's
in-progress typing: `Host::commands` is synchronous, so the snapshot and the
restore bracket **only the frame's own writes**. There is no "in between" for a
keystroke to land in.

Injected fault: dropping `self.view_state = snapshot.view_state` from
`App::restore` fails this test and only this test.

### a_refused_mutation_rolls_back_its_predecessor_and_the_abort_flag

The `Rejection::MutationRefused` arm — the one the first hand-back wrongly called
unreachable — plus the batch-outcome flags going back with everything else.

**`batch_aborted` was left `true` after a rolled-back frame.** Inert at the time,
since every `Command::Model` resets it at the top of `apply_effects`, but it is
the same shape as the stale-flag bug two commits earlier and `App::restore` is its
natural home. `App`'s own doc comment names four more "last batch" fields and
says `last_reshaped_views` is "the shape `batch_aborted` takes"; `restore` was
clearing three of the four and neither snapshotting nor restoring any of them.
All six now round-trip through `AppSnapshot`, which is why the type was renamed
from `DocumentSnapshot` — it is no longer only the document.

**The fixture detail that matters**: a *duplicate* `AddCategory` is idempotent, so
the obvious way to reach the refusal proves nothing. This fills to
`MAX_CATEGORIES` (12) and overflows. And the predecessor is an `AddItem` rather
than a `SetCell`, because adding eleven categories reshapes the view enough that
a record-addressed predecessor is refused `RecordGone` **first** — hiding the arm
under test behind an earlier one.

Injected fault: dropping `self.batch_aborted = snapshot.batch_aborted` fails this
test and only this test.

### a_frame_whose_write_would_land_nowhere_is_refused_rather_than_dropped

**Found by reviewing improvise-yqpd's own work, and it was a hole inside the fix
for the hole.** The first version of `reduce::broken_chain` carried a
create-exemption — the same `will_be_created_by_the_frame` predicate the pre-frame
check uses — on the reasoning that the two must not drift. The hand-back recorded
that arm as "correctness-preserving but unobservable", because removing the filter
failed no test.

That reasoning was wrong, and the fixture that shows it is the one the hand-back
said did not exist. `[AddItem("West"), HideItem("West"), SetCell(West)]` observed
**`rejected=false landed=false`**: the frame reported accepted, with the write
silently gone — *precisely* the partial-and-silent failure clause 4 exists to
eliminate, surviving inside clause 4's own implementation.

**The argument that settles it.** `App::resolve_cell` is gated by
`resolve_record`, so `resolve_record(hash).is_none()` is exactly the condition
under which `model_effects` returns an empty vec and the write vanishes. And
`SetCell`/`ClearCell` are the only variants `records_named` names, so there is no
command for which a non-resolving record still does something useful. Refusing is
therefore strictly better than dropping, everywhere.

**Scope: this closes the record-addressed INSTANCE, not the class.**
`broken_chain` asks `records_named`, which names only `SetCell`/`ClearCell`, so a
frame can still apply partially and silently through a command that names no
record. Two measured as accepted: `[DeleteView("Scratch"),
SetNumberFormat{view:"Scratch"}]`, where the effect runs and no-ops inside
`apply`; and `[RemoveCategory("Region"), AddItem{category:"Region"}]`, where the
effect *fails* and writes its error to the **host's** status line without
aborting, so the client is never told. Closing the class needs effects to report
failure rather than narrate it — roughly fifty of them — which is a different
change from clause 4 and is stated here rather than implied.

**The exemption belongs at CHECK time and nowhere else**, and that asymmetry is
the point: before the frame runs, a record the frame will create legitimately does
not exist, so `[AddItem(East), SetCell(East)]` must pass — `kb.pl`'s
`the_create_exemption_is_a_property_not_a_list` is unchanged and still describes
`stale_records_in_a_frame`. At apply time the record has either been created or it
has not, and "has not" means the write is lost.

Injected fault: restoring any exemption to `broken_chain` fails this test, plus
`a_frame_whose_later_mutation_fails_is_undone_whole` and the mirror test.

### a_frame_of_view_commands_is_not_refused_by_an_earlier_batchs_abort

**The second review finding, and a live bug improvise-yqpd introduced.**
`App::batch_aborted` is reset *only* inside `apply_effects` (gotcha 21), and
`Host::commands` read `last_batch_aborted()` after **every** command in a frame.
A command that runs no effects — every view command, and `Command::SwitchView`,
which `apply_checked` routes to `apply_switch` rather than through
`apply_command` — leaves the flag exactly as some earlier batch left it.

So a TUI `:w /bad/path` raised the flag, and the next frame a connected browser
sent was refused `MutationRefused` and rolled back, having done nothing wrong.
Reproduced by aborting a pathless `effect::Save` directly on the `App` and then
sending `[SetSelected]`.

**Two arms, because frame-level gating alone is not the fix.** The first arm is a
pure view frame, which the `Command::Model`-presence gate in `commands` handles.
The second is the **mixed** frame `[SetSelected, SetCell]`, which that gate lets
through and which only the per-command `mutation &&` guard answers — the view
command runs no effects, so the flag under it is still the stale one. Injecting
`if { let _ = mutation; self.app.last_batch_aborted() }` fails this test and only
this test, which is what shows the per-command guard is load-bearing rather than
belt-and-braces.

The same `Command::Model`-presence predicate is why a frame of pure cursor keys
now pays for neither the `Workbook` snapshot nor the `Sessions` clone — the
browser's common path, one frame per keypress.

### the_mirror_settles_its_mode_against_every_grid_it_is_later_projected

improvise-el9h. `ClientSession::fold` settled `view.mode` against `self.grid` — the
grid that session was **last projected** — and `project_diff` then assigned
`self.grid = desired` and never settled again. The client does not have that shape:
`ClientState::fold` calls `settle_mode()` on every fold including
`Command::Project`, so when the new view's grid arrives the client re-settles and
the mirror does not.

`[SwitchView(records-shaped view), ChangeMode(Normal)]` is the shape — precisely the
case PR #98's settle exists to serve. **Demonstrated before the fix**: the client
ended on `RecordsNormal` and the mirror on `Normal`. The fix is `settle_mode` called
from `fold` *and* at the end of `project_diff` once the grid has moved, which also
closes the three bypass doors (`apply_switch`, `adopt_deleted_views`,
`adopt_cursor_epoch` write the mirror with bare `reduce_view`): whatever they leave,
the next projection settles.

**Two preconditions, and one of them caught a broken fixture on its first run.** The
test asserts the client starts on `Normal`, so the settle has somewhere to move the
mode *from*. And `app_with_a_records_shaped_second_view` asserts the second view
really reaches records mode — which failed immediately, because the view was built
with `View::new` and a fresh `View` carries no axis entry for `_Index`/`_Dim`, so
`is_records_mode` can never fire on one. Without that assertion the test would have
compared two modes that were both `Normal` for a reason having nothing to do with
the bug, and passed against the unfixed code.

The yardstick is a real `improvise_protocol::ClientState` fed the same projections
in arrival order, not a hand-written expectation: the property is that the mirror
and the client run the same function over the same stream.

### a_failed_frame_puts_the_askers_mirror_back_on_the_view_it_was_on

Clause 4's least obvious cost, in the ticket's own words: **rolling back the `App`
is not enough.** `Host::command` projects per command and `apply_switch` /
`adopt_deleted_views` / `adopt_cursor_epoch` move mirrors inline, so by the time a
later command fails, every session's mirror has already moved. `Host::commands`
therefore snapshots the whole `Sessions` beside the document and restores both.

**The first version of this test was vacuous, and the injection is what showed it.**
It seated a bystander on a second view and asserted the bystander was told nothing
and its mirror was unmoved — both of which hold when `undo_frame` restores *nothing
at all*, because the frame's projections are discarded from `out` either way and
nothing ever moved that bystander. It answered *did the frame return early* where it
meant *was the mirror put back*.

The rewrite aims at the one mirror move a failed frame really makes and that
**nothing in the stream can ever correct**: the asker's own `SwitchView`. An
accepted switch is relayed to nobody (gotcha 36), so a mirror left on the switched-to
view diverges from a client that was told the frame was refused, permanently and
silently. The fixture asserts up front that the mirror starts on `Default` — the
direction the default cannot supply — so the closing assertion is not satisfied by a
mirror nothing ever wrote. Removing `self.sessions = mirrors` while keeping the
document restore fails this test and only this test.

### a_frame_that_creates_the_record_it_writes_is_not_a_broken_chain

The other direction: clause 4 must not fire on `[AddItem("West"), SetCell(West)]`,
which is the exemption ADR 001 states as a **property** — a record that neither
resolves nor is claimed is one the frame is creating.

**Be clear about what this does and does not catch.** It catches `broken_chain`
being asked against the *pre-frame* document instead of the running one, which is
the plausible implementation error (passing the snapshot where the live `App`
belongs) and would refuse every data-dependent chain. It does **not** catch the
exemption filter being dropped, for the reason recorded above: `resolve_record`
falls back to the layout, so `West` already resolves once `AddItem` has run. A
fixture that discriminated the filter would need a created record that never enters
the layout; none exists in this workspace and one was not invented for it.

#### Helper — `frame_basis`

Not a test — the frame-level claim that test and the 955n tests state.

The basis a caught-up client would state for a **frame** naming `regions`: the two
document tiers plus each named record's current counter.

`reduce::caught_up` cannot stand in for this, and the difference is exactly what a
frame is: it builds the claim for **one command**, so a frame's records beyond the
first are simply absent from it and come back `RecordUnstated`. A real client calls
`BasisCache::claim` once with every record its frame names, which is what this
reproduces.

**Name only records that exist.** `claim` *omits* a record the cache has never
heard of; this helper always includes one, so pointed at a record that is not there
it fabricates `(hash, 0)` — a claim no real client could compose, and precisely the
input that flips the create exemption in `CheckedBatch`. For a record the frame is
*creating*, leave it out, which is what a client would do.

### an_accepted_switch_is_relayed_to_nobody

The successor to `a_bystander_is_sent_the_broadcast_and_the_asker_is_not`, which
pinned the pre-improvise-jytm relay (Copilot, PR #10). An accepted switch is now
owed to nobody: the asker folded it, and no other client's view moves.

It asserts the bystander gets **no frame at all**, not merely no `SwitchView`. A
relay that vanished while the projection fan-out still repainted every client would
pass the narrower check and still be pushing one client's view change at the rest.

### a_geometry_command_relays_the_cursor_reset_to_the_asker_too

The other audience, and the reason `Broadcast` carries one rather than the relay
deciding. A geometry change invalidates every cursor into the document
(improvise-gvnb), and the asker's is not exempt: a `Command::Model` is the one kind
a client cannot fold optimistically, so nothing has reset it locally and skipping it
would leave the very client that asked positioned in a grid it no longer has.

### a_stale_batch_is_refused_to_the_asker_alone

A refused batch is answered to the **asker and nobody else** (ADR 001). Nothing
happened, so there is nothing to tell the others — and the asker is told only which
class of staleness it was, because it already holds what is current.

### folding_the_relay_puts_a_refused_asker_back

What `Broadcast::To` is *for*, asserted end to end through `Host` rather than
through `Sessions`: a client that folded a view name the document refused is put
back by the frame the host sends it, so the cursor its next grid is positioned by is
read under a key the document has (gotcha 24).

The client slice is folded **before** the command is handed to the host, which is
what makes the assertion mean anything: a test that never folded would see
"Default" whatever the relay did. The bystander is asserted silent in the same test,
because a `ToAll` correction would also put the asker back and would be wrong.

### a_document_swap_sends_every_client_a_fresh_snapshot

AC of gotcha 39: a command that replaced the document owes *every* client a fresh
`Snapshot`, not a diff — including the client that did not ask.

#### Helper — `host_over_a_loadable_document`

Not a test — the host that test and `a_load_that_fails_is_not_a_rebase` drive.

A host whose document can be replaced out from under its clients: a second document
sits in the storage backend under a name `LoadModel` can reach.

### a_client_can_ask_its_host_to_save

improvise-368h. **The bug**: `effect::Save` is `EffectKind::ViewOnly` — correctly,
since it writes *this host's* path — so the derivation that builds `ModelCommand`
out of the `EffectKind::Model` effects could not reach it (gotcha 38). There was no
save on the wire at all, so **no client on any transport could cause its host to
write**: a websocket session edited the document in memory and the file on disk
never moved unless a human was at a TUI that was not there.

The derivation was not wrong; it was complete over the mutations and saving is not
one. So the variant is *added* rather than derived, and it carries no path: a client
cannot see the host's storage, and which document a host is hosting is the host's
own fact.

### a_save_with_nowhere_to_write_is_refused_and_told_only_to_the_asker

improvise-368h: the "what if there is no path" half. Not a panic and not a silent
no-op — the two answers the ticket ruled out — but the vocabulary that already
exists for "nothing happened and here is why", addressed to the asker alone
(gotcha 39).

The second client is not decoration: a refusal that leaked to the bystanders would
be telling them about a failure of somebody else's host-local request, and this
asserts it does not.

### a_write_that_fails_is_refused_rather_than_reported_as_saved

improvise-368h: a write that was attempted and failed must not be reported as a
success. `Save` signals failure by aborting its batch (gotcha 21), and the batch
here is exactly `[Save]`, so an abort is that failure and nothing else.

The fixture fails for a reason that is real rather than mocked: a `PhysicalFs`
pointed at a directory that does not exist. It creates nothing, so the test leaves
no trace on disk.

### every_frame_survives_a_json_round_trip

The wire is JSON to start with (the ticket's choice, kept), so both directions must
survive a round trip through it — including the one point at which `ViewState` is
serialized (gotcha 25).

The inbound half carries a payload-bearing variant on each side of the envelope
(`doctrine(populate_before_roundtrip)`): a `Viewport` and a `Command` carrying a
`CellValue::Text`, which is the arm the outbound half cannot reach — a text cell
aggregates to nothing under a None-axis category, so it never becomes a
`CellDisplay`.

*Note:* `host.rs` also carries a second `a_load_that_fails_is_not_a_rebase`
(`doctrine(check_the_checker)`: inject the fault the test targets — a load that does
not happen — and the snapshot must not appear; a test that saw `Snapshot` on every
command would pass `a_document_swap_sends_every_client_a_fresh_snapshot`). It shares
its name with the `ui/project.rs` test of the same name, which is why there is one
heading for both above.

---

## `ui/cat_tree.rs`

### the_tree_carries_the_virtual_categories_and_measure_includes_formula_targets

improvise-vjms: the panel lists the virtual categories alongside the regular ones,
and `_Measure`'s items are *derived* — explicit items ∪ formula targets (gotcha 7).
An implementation reading `category("_Measure").items` passes every other assertion
in the module and silently drops `Profit`, which is the failure improvise-ete was
closed for proposing.

---

## `command/keymap.rs`

### f1_and_question_mark_both_open_help

improvise-th7m: `F1` was bound twice on consecutive lines — first to a bare
`enter-mode` with no args, then to `enter-mode ["help"]`. The first was dead,
because `Keymap::bind` *inserts* and the later binding wins; had it ever been the
surviving one it would have failed `require_args` at dispatch and made `F1` do
nothing.

The risk in deleting a line of a duplicated pair is deleting the wrong one, so this
pins which binding `F1` is meant to carry: the same one `?` carries, which is what
identified it as the intended survivor.

### every_bound_command_name_is_registered

improvise-9cn: `Keymap::dispatch` resolves a binding through
`registry.interactive(name, args, ctx).ok()?`, so a binding is a silent no-op key in
two ways, not one: the name may not be registered, *or* the args may not parse. Both
walks live in this test.

The args half matters because improvise-s13 moved bindings out of the nullary class
the name check covers: `enter-mode ["export-prompt"]` and
`enter-mode ["formula-edit"]` carry a mode name that `parse_mode_name` must accept,
and a typo like `"export_prompt"` compiles, registers, and produces a dead
`Ctrl-e`. (Gotcha 15.)

It builds the context in the mode whose keymap it is walking, because a few parsers
read `ctx.mode` — `commit-buffer item` resolves its target category off
`AppMode::ItemAdd`.

improvise-piyp. This entry is about the test's **reach**, not its logic — the logic was right
and demonstrably so, which is what made the gap hard to see.

It had two blind spots, both found by injection rather than by reading:

- the walk matched `Some(Binding::Prefix(_)) | None => {}`, so it **never
  descended into a prefix sub-keymap**. Everything under `g`, `y` and `Z` was
  unchecked, in every mode — and `RecordsNormal` reaches Normal's three
  through `Keymaps::prefix_keymap`, so it was unchecked twice over.
- its key space was a **second copy** of `tests_support::probe_keys_pub`,
  identical but for the two F-keys the public one carries. So a name bound on
  any function key was invisible here while `dump_every_binding` probed it.

Injected at `523d571`: one unregistered name under the `g` prefix and one on
`F(5)` in Normal. `cargo test --workspace` came back **1595 passed, 0 failed**
— byte-identical to the clean baseline. The positive control that separates
"untested path" from "broken test" is the same bogus name on `Char('~')`, a
probed non-prefix key: that fails. The guard worked; it could not see there.

Both are closed here. The walk is `reachable_names`, which recurses into
`Binding::Prefix` — terminating because `prefix_keymaps_are_flat` says the
nesting is one deep — and `probe_keys` now **delegates** to `probe_keys_pub`
rather than restating it, which is the part that matters: two lists of "the
keys a test probes" is the same disease as two keymap spines, and the F-key
gap was exactly the drift between them. `probe_keys_pub` covers `F(1)..=F(12)`
rather than the two that happen to be bound, since the point is to see a
binding nobody expected.

Re-running the two injections against the fixed test: the prefix one fails
with `Normal binds "no-such-command-under-a-prefix", which no registry entry
provides`, and the F-key one with the matching message for `F(5)`.

Worth noting for contrast, because reading either test alone misleads:
`nothing_declined_is_also_composed` in `resolve.rs` already recursed into
prefixes and into parents. The **resolution** side was guarded where the
**registration** side was not, and only comparing the two shows which half is
real.

### prefix_keymaps_are_flat

improvise-jc8: an armed prefix is stored as the key that opened it and resolved
against the CURRENT MODE's keymap. That is only correct while prefixes are flat — a
`Prefix` binding nested inside a sub-keymap would be resolved against the mode map,
where it does not exist, and the second half of the sequence would silently do
nothing. Adding a nested prefix must fail here rather than in the terminal.

### every_panel_keymap_shares_the_same_spine

improvise-4pg: the three panel keymaps were written out separately, so a spine
binding could be changed in one and forgotten in the others. They share
`PanelKeymap` now; this pins what "the spine" is, including that each panel's cursor
keys move its *own* cursor. (Gotcha 12.)

### every_mode_with_panel_toggles_binds_all_of_them

Every panel's toggle key is bound in Normal and in all three panel modes, so focus
can move between any two and the focused one closes. A fourth panel gets all four
keymaps from one `Panel::ALL` entry.

### panel_toggle_keys_dispatch_to_real_effects

`Panel::name()` has to spell what the registry's panel parser accepts. A lookup-only
assertion would not catch a mismatch — `dispatch` swallows an unresolvable command
with `.ok()?` — so the toggle key must come back with real effects.

### every_text_entry_mode_types_into_the_buffer_its_prompt_names

improvise-k8h: the buffer name used to be written on both sides of the crate —
`AppMode`'s `MinibufferConfig` names the buffer the prompt describes, and the keymap
names the buffer `append-char` edits. If those disagree the mode prompts over one
buffer while every keystroke edits another, with nothing to notice. `TextEntry` now
reads the name off the mode, and this pins the result for every text-entry mode.
(Gotcha 11.)

#### Helper — `text_entry_modes_with_exits`

Not a test — the list that test and the two below it walk.

The seven modes that carry a `MinibufferConfig`, each with the keys that commit or
leave it. `SearchMode` is not among them: it is `Normal` plus a flag, so it has no
`AppMode` variant and no minibuffer config to agree with.

### every_typing_keymap_belongs_to_a_text_entry_mode

The list in `text_entry_modes_with_exits` must stay complete: every keymap that
types into a buffer belongs to a mode whose minibuffer names that same buffer. Only
`SearchMode`, which has no `AppMode` variant, is exempt.

### text_entry_modes_clear_the_buffer_they_type_into

Every minibuffer mode types into exactly one buffer, and every key that commits or
leaves must clear that same buffer on the way — a buffer left behind pre-fills the
next entry with the last one. `SearchMode` is excluded on purpose: it has no commit
and its query is owned by `exit-search-mode`, not by `clear-buffer`.

The buffer is derived from each mode rather than restated in the test —
`every_text_entry_mode_types_into_the_buffer_its_prompt_names` owns pinning what it
should be (improvise-k8h).

### minibuffer_enter_dispatches_to_a_registered_commit

improvise-6t3: the three commit commands collapsed into a single
`commit-buffer <kind>`. `dispatch` swallows an unresolvable command name with
`.ok()?`, so a lookup-only assertion would still pass had the rename missed the
registry — Enter has to come back with real effects.

### dump_every_binding

The instrument `doctrine(binding_dump_diff)` asks for: every `(mode, key, modifier)`
lookup across all the modes, written to the file named by `IMPROVISE_BINDING_DUMP`.

Most bindings have no direct test, so a green suite is not evidence that a keymap
refactor preserved them. Dump before, dump after, diff. It is inert without the
variable: a test that writes a file nobody asked for is a surprise, not a check.

This is what proved the keymap move to `improvise-protocol` changed nothing — 4021
resolved lookups, byte-identical either side — and the diff was shown to be
discriminating by renaming one binding and watching eight lines move.

---

## `command/cmd/commit.rs`


Not an assertion — `doctrine(binding_dump_diff)`'s instrument, and the
evidence improvise-636d actually rests on, because a green suite is not
evidence for a keymap refactor.

Diffed across the global change: **84 added lookups, 0 removed, 0 changed**.
No binding anywhere answers differently than it did, which is the property
that matters for a change that moves bindings between tables. The additions
are 16 rows of `?` reaching the three panels and TileSelect, 1 of Help's new
`?`, 24 of the panel toggles reaching Help and TileSelect, and 40 that are not
a binding change at all — the widened `probe_keys_pub` hitting ImportWizard's
`KeyPattern::Any` on `F(3)..=F(12)`.

**`:` produced zero rows**, which is the result worth keeping: every mode that
now inherits it already bound it locally, so moving it into the global changed
no answer anywhere. That is what "the duplication was real and consistent"
looks like from the outside — and it is also why nothing in the suite could
have caught the two spines drifting apart, had they ever done so.

### invalid_formula_reports_the_parse_error_not_success

improvise-6t3: the commit pushed `set_status("Formula added")` *after* `AddFormula`,
so the parse error that `AddFormula::apply` writes to the status bar was immediately
overwritten. An unparseable formula was dropped while the UI reported success.

The add effect is the only party that knows whether it worked, so it has to get the
last word on the status line.

### commit_and_advance_threads_edit_mode_to_enter_edit_at_cursor

`CommitAndAdvance` must thread its `edit_mode` through to the trailing
`EnterEditAtCursor` effect so the post-commit re-edit lands in the mode the keymap
requested. The command never reads `ctx.mode`.

### commit_save_writes_the_model_to_the_typed_path

improvise-lae: the save prompt has to commit to `SaveAs` and not to `ExportCsv`,
which is what makes it a separate mode from `ExportPrompt` rather than a second
binding onto it — `ExportCsv` would write CSV under whatever `.improv` name the user
typed.

`SaveAs` and not `Save`, because `Save` writes the path the document already has,
and the whole reason this prompt exists is that there isn't one.

### commit_save_on_an_empty_path_writes_nothing

An empty prompt must not reach `SaveAs`: `SaveAs(PathBuf::from(""))` asks the
backend to write to `""`, and the user then reads whatever I/O error that produces
instead of "you typed nothing".

### commit_export_bails_on_an_empty_path

improvise-z05r's first arm. `CommitExport` did not bail on an empty path where
`CommitBuffer`'s three kinds all do (gotcha 20: buffer to trim to bail-on-empty to
dirty to status is the shared spine), so Enter on an empty export prompt asked the
backend to write to `""` and the user read whatever I/O error that produced.

The same change reorders the batch: the mode change now comes **first**, because
`ExportCsv` aborts on a failed write as of z05r's other arm, and an abort at the tail
would strand the user in `ExportPrompt` — where `draw_bottom_bar` renders the
minibuffer *instead of* the status line, so the export error is never drawn at all
and the operation appears to do nothing (gotcha 21). That ordering is asserted in
`commit_export_produces_export_and_normal_mode`, by index and not by presence.


---

## `command/cmd/fit.rs`

### the_command_installs_a_formula_the_document_evaluates

The end-to-end claim of improvise-vx3, asserted against a real document: a command
fits, the formula lands through the ordinary `add_formula` path, and the new measure
evaluates to numbers.

#### Helper — `workbook`

Not a test — the document every `fit.rs` test fits against.

A document a tree can learn exactly — `Score` steps at `Cost` 50 — in the units a
real document uses.

**Every category is a `Label`**, so the only feature is the numeric measure `Cost`
and the champion cannot reach an unresolvable one. An earlier version left `Region`
regular with one item and claimed `simplify` would prune the constant one-hot; it
does not, because `build_categorical` records `(0.0, 1.0)` for every one-hot
whatever the column holds — so these tests were passing on seed luck (Copilot on
PR #25). See `improvise_gp::fit`'s own fixture for the same correction.

### the_status_line_reports_quality_against_the_baseline

improvise-vx3 is emphatic that the rule alone is not the deliverable: a tidy tree
that generalizes no better than the mean is worse than useless. So the status line
carries the held-out comparison, what the rule leans on, and the seed that
reproduces it.

### a_refusal_names_what_went_wrong_and_stores_nothing

A refusal reaches the user *whole*. "fit failed" is not actionable; naming the
measure that does not exist is (improvise-8xhy's argument, arriving at the place a
user meets it).

### the_seed_argument_changes_the_search

Gotcha 37b: the seed is half of what reproduces a fit, so it is an argument — and a
different one really does search differently.

It asserts on the **rule**, never on the batch's debug string. The first version
compared `effects_debug` over the whole vec, which was vacuous: `status_line` prints
`seed {:#x}`, so the two strings differ because the seed was *printed*, and
`fit_measure` could have ignored `request.seed` entirely with the test still green
(`doctrine(check_the_checker)`).

`AddFormula` has exactly two fields — the formula text and the target category — and
the seed reaches neither, so its `Debug` cannot differ for the reason the batch's
did. (A substring guard was tried and is wrong: the seeds `1` and `99` occur inside
perfectly ordinary thresholds.)

### the_seed_the_status_line_prints_is_a_seed_this_command_accepts

Copilot on PR #25. The status line advertises the seed as `{:#x}` and tells the user
that is what reproduces the fit, so the one instruction it gives has to work on the
value it just showed them — a decimal-only parser made copying it back an error.

Asserted as a *round trip through the status line*, not against a hand-written
`"0x…"`: a literal would still pass if `status_line` changed how it spells the
number, which is the half that breaks.

---

## `command/cmd/grid.rs`

### drill_into_formula_cell_returns_data_records

Drilling into a formula cell (`Profit = Revenue - Cost`) must return the underlying
data records, not an empty result set. The formula target coordinate is stripped
from the drill key so that `matching_cells` finds the raw data backing the formula
(gotcha 10).

### drilling_fixes_its_coordinates_on_filter_not_page

improvise-dz8: a drilled coordinate is a constraint of the drill, not a dimension
the user may page through, so it lands on `Filter`. It used to land on `Page`, which
put it in the tile bar and under `[`/`]` as if the user could move it (gotcha 27).

---

## `command/cmd/mode.rs`

### enter_mode_export_prompt_sets_mode_and_clears_its_buffer

The `enter-export-prompt` shim is gone (improvise-s13); the generic `enter-mode`
carries the mode name instead, and clears the prompt's buffer on the way in.

### enter_mode_clears_a_search_keyed_minibuffer_through_write_buffer

improvise-gdl made `write_buffer` the single accessor for a runtime-named buffer,
because "search" is not an entry in the buffers map but its own `ViewState` field.
`EnterMode` built a `SetBuffer` directly, so a mode whose `buffer_key` is "search"
would clear a map entry nobody reads and leave the previous query on screen — the
exact desync gdl closed. No shipped `AppMode` names that buffer yet, so the mode here
is constructed by hand; that is what makes the test a demonstration rather than a
restatement.

### edit_or_drill_passes_records_editing_mode_through

`EditOrDrill` must trust its `edit_mode` parameter rather than checking `ctx.mode` —
the records-normal keymap supplies `records-editing`, but the command itself never
inspects the runtime mode. This is the parallel of the deleted
`enter_edit_mode_produces_editing_mode` test for the records branch.

### enter_edit_at_cursor_cmd_passes_target_mode_to_effect

`EnterEditAtCursorCmd` must hand its `target_mode` straight through to the
`EnterEditAtCursor` effect — the keymap (records `o` sequence or
commit-and-advance) decides; the command never inspects `ctx`. It fronts the pair
with `RebuildLayout` so the cursor is read against a layout that includes whatever
ran earlier in the sequence (improvise-ehz, gotcha 23).

---

## `command/cmd/navigation.rs`

### page_cycling_skips_filter_coordinates

improvise-dz8: `[` and `]` page through Page-axis categories. A coordinate the drill
fixed sits on `Filter` and must be invisible to them, or the user could page out of
the drill they are inside.

### enter_advance_at_bottom_right_emits_abort_chain

At bottom-right `EnterAdvance` has no place to go, so it emits a single `AbortChain`
effect. Trailing effects in a `CommitAndAdvance` batch (`EnterEditAtCursor`) are
then skipped, which is how "Enter at bottom-right commits and exits editing" is
realised.

---

## `command/cmd/panel.rs`

### cycle_panel_focus_advances_from_formula_to_category

Bug improvise-dqn: `CyclePanelFocus` always picked the first open panel, ignoring
the currently focused panel. With formula + category open and focus on the formula
panel, cycling must move to the category panel.

### create_and_switch_view_names_incrementally

improvise-aaw asked for this command to be replaced by a keymap
`Binding::Sequence [create-view X, switch-view X, enter-mode normal]`. It cannot be:
a `Sequence` carries **literal** args, and the name here is computed from `ctx` — so
a Sequence would hard-code one name and the second `n` in the view panel would
collide with the first.

The test was called `..._names_incrementally` and asserted only that
`CreateView`/`SwitchView`/`Normal` appeared — it never looked at the name, which is
the one thing that makes this command a command and not a composition. Now it does,
at two different view counts, because a single count cannot tell "derived from the
workbook" from "constant".

---

## `command/cmd/registry.rs`

### every_entry_constructs_a_command_with_its_registered_name

improvise-9cn / improvise-61f: every registration derives its name from a prototype
instance, so a typo in the name is impossible. What the types do *not* catch is the
commands whose `name()` branches on a field — `ViewNavigate`, `TileAxisOp`,
`ToggleGroupAtCursor`, `CommitAndAdvance`. Their prototype and the value their
constructor builds are the same type, so registering `ViewNavigate { forward: true }`
while building `forward: false` compiles cleanly and files the command under the
wrong name, where dispatch silently misses. Constructing every entry both ways and
comparing names is the only check that reaches that class. (Gotcha 15.)

Some entries are deliberately one-path — `commit-and-advance-right` reads the edit
buffer and cursor, so its parse fn always errors — but every entry must be reachable
by at least one path, or its name goes unchecked.

### registered_names_are_unique

A variant mix-up does not show up as a name mismatch — it shows up as a **collision**.
Registering `ViewNavigate { forward: false }` where `forward: true` belongs files a
second entry under "view-back" and removes "view-forward" from the registry
entirely; `parse` takes the first match, so every surviving entry still agrees with
its own name. Uniqueness is what catches it.

### the_context_reports_each_panels_own_open_flag

improvise-y4xg and improvise-qsmp, two shapes of one defect. `registry.rs` mapped
`Panel` to a `CmdContext` field in two places, and `App::cmd_context` copies the
three flags out of `ViewState` in a third. **Both were demonstrated unguarded**:
swapping Formula and Category at the two registry sites left the workspace at 1400
passing, and hardcoding the `category_panel_open` propagator to `true` left it at
1394.

Nothing exercised either, because `cmd/mod.rs::test_helpers` sets the panel flags
explicitly on the context it builds — the command tests construct the value they then
assert on, so no test ever drove the real propagator.

The registry mapping is gone: both sites call `CmdContext::panel_open(panel)`, the
one mapping on this side of the wall. That leaves two permutable sites, and the test
is aimed at both from a path neither controls — it writes `view_state`'s three fields
**by name** and reads back through the enum accessor on a context built by the real
`App::cmd_context`. PR #70's review caught the weaker shape of this: a panel test that
zeroed and read through the same four accessors passed a consistent permutation.

Injected fault: swap `Panel::Formula` and `Panel::Category` in `panel_open`. Fails.
Injected fault: hardcode the `category_panel_open` propagator to `true`. Fails.

### toggle_panel_visibility_reads_the_panel_it_names

The behavioural half, and the one that reaches the registry. Its interactive arm
passes `currently_open` straight through, so with exactly one panel open the command
built for each of the three names carries its own panel's flag. Both injections above
fail here too.

### a_focused_open_panel_closes_and_a_closed_one_opens

improvise-qsmp's own scenario: `toggle-panel-and-focus` decides close only when the
panel is *both* open and focused, so the case that separates a right propagator from
a stuck-true one needs the mode set to the panel's own. With the flag stuck true,
pressing a panel's key on a closed panel decides to *close* it and the panel becomes
unopenable by its own key — which is what this test's second arm names.

### Helper — `open_only`

Not a test. It spells `formula_panel_open`, `category_panel_open` and
`view_panel_open` as fields rather than looping over `Panel::ALL` through an
accessor, and that is the whole evidentiary value: a permutation in any mapping
between the `Panel` enum and those three booleans has to disagree with this helper.


---

## `command/cmd/text_buffer.rs`

### writing_the_search_buffer_targets_the_search_query

Writing the "search" buffer means writing the search query, and `write_buffer` is
the one place that knows it (improvise-gdl). Every buffer-writing command routes
through it, so `clear-buffer search` clears the query rather than parking an unread
entry in the map.

### colon_wq_saves_then_quits

improvise-git proposed deleting the `SaveAndQuit` command and binding `ZZ` to a
`Binding::Sequence [w, force-quit]` instead — the repo's "do not name a command
AAndB" principle.

That premise is incomplete, and this test is why. `wq` is not only a keybinding: it
is a **registered command name**, so `:wq` reaches it through the registry, and
`Quit`'s own unsaved-changes message advertises it ("Use :q! to force quit or :wq to
save+quit"). A keymap `Sequence` is not reachable from the `:` line, so deleting the
command deletes the ex-command with it.

Keeping the command *and* adding the Sequence would leave two independently-written
spellings of "save then quit", which is the duplication this codebase keeps curing
rather than creating. So this pins the surface, and the conversion is declined on it.

### colon_enter_edit_at_cursor_is_not_snapped_back_to_normal

improvise-6sq: `ExecuteCommand` appends `change_mode(Normal)` unless some effect
reports `changes_mode()`. `EnterEditAtCursor` sets `view_state.mode` but did not
override it, so `:enter-edit-at-cursor editing` produced
`[RebuildLayout, EnterEditAtCursor(Editing), ChangeMode(Normal)]` — the trailing
snap-back undid the edit mode and the user never got a cursor in the minibuffer.

Asserting on `changes_mode()` alone is not enough here: the appended
`ChangeMode(Normal)` reports `true` itself, so the assertion has to name the
snap-back it is ruling out.

### colon_import_stays_in_the_wizard_on_success

improvise-toc: `StartImportWizard` swaps the mode only when the file parses, so a
per-type `changes_mode()` cannot describe it. Reporting `false` made
`ExecuteCommand` append a return-to-Normal on top of the success path:
`:import d.json` opened the wizard and was immediately snapped back, leaving the
wizard drawn under the Normal keymap. (Gotcha 31.)

### colon_import_falls_back_to_normal_on_failure

The other half of the same bug: reporting `true` unconditionally would strand the
user in `CommandMode` when the import fails. A failed `:import` must still fall back
to Normal.

### colon_view_back_respects_whether_history_exists

The `ViewBack` half of improvise-toc: it restores the stored frame's mode only when
the back-stack is non-empty. With history present, `:view-back` must keep the mode it
restored; with an empty stack it falls through to Normal like any other no-op
command.

---

## `command/cmd/wizard.rs`

### name_model_enter_builds_the_model

Enter on the name step builds the workbook into the wizard rather than advancing the
step machine; every other printable key edits the name (gotcha 22).

### done_confirms_on_enter_and_steps_back_on_esc

On the confirmation step Enter installs the built workbook and Esc steps back — it
does not throw the import away, and no key closes the wizard outright
(improvise-907, from improvise-c6j).

### a_formula_owned_cell_is_projected_as_derived_and_its_siblings_are_not

improvise-ir7l. `project::cell_id` decided `Record` vs `Aggregate` purely on
whether a regular category sat on `Axis::None`; it never asked whether the
`_Measure` coordinate named a formula target. So a formula cell was a `Record`,
`render.rs::editable` said true, and a write to it was accepted, stored and
then shadowed forever by `formula_cache`.

The fault injected: replacing the `derived` lookup with `None`, which is the
pre-fix code exactly. The test then reports `Record(RecordCell { measure:
Some("Profit") })` where it wants `Derived { target: "Profit" }`. The siblings
are asserted in the same test so a producer that answered `Derived` for *every*
cell could not pass.

### an_aggregating_formula_cell_is_still_projected_as_an_aggregate

The negative arm of the same change, and the one a careless fix breaks: gotcha
10 says drilling into a formula cell is supported, so `Aggregate` must still
win where the cell aggregates. Proved discriminating by returning `derived`
before the aggregate branch — the cell then reads `Derived` and becomes
undrillable, with nothing else in the suite noticing.

### edit_or_drill_refuses_a_formula_owned_cell

The TUI half. `EditOrDrill` never asked the question either, so the terminal
offered the same editor over the same shadowed cell. Landing only the
projection half would have made the two front ends disagree, which is the
defect design-principles §5 is about. Fault injected by deleting the guard: the
command returns `[SetBuffer{edit}, ChangeMode(Editing)]` again.

### a_save_clears_the_dirty_marker_on_a_client_although_nothing_is_dirty

improvise-z754, and the arm the whole design turns on. `effect::Save` is
`EffectKind::ViewOnly`, so `Dirty::of` reports `NOTHING` — the test asserts
that first, so it cannot quietly stop being the interesting case if `Save` is
ever reclassified. A title bar gated on `Dirty` would therefore leave every
client showing `[+]` for a document that has just been saved, forever, with
nothing in the stream to correct it.

Fault injected: wrapping `document_update` in `if !dirty.is_empty()`, which is
the obvious implementation and the wrong one. The test fails on the cleared
arm. `a_title_bar_that_did_not_move_is_not_projected` is its counterweight —
without it, "project the title bar every batch" would pass this one.

### the_title_bar_names_the_document_without_naming_the_path

The fixture's path is `/home/someone/private/books.improv` on purpose: a test
whose fixture is a bare file name cannot tell `file_name()` from the whole
path, and the failure mode is a socket server handing a remote client the
host's directory layout.

### a_creating_column_is_named_rather_than_left_as_no_cell

improvise-poia. Both `+` and `+cat` projected `None`, which the protocol
defines as "no projected cell" — the same conflation gotcha 42 names for
aggregates, arriving a third time.

The test walks **every** column of a records grid and counts, rather than
asserting the two creating columns alone. That matters in the direction a
weaker test fails: a producer that answered `Create` for everything, or that
made the `Region` coordinate column nameable, would pass a two-position check
and fails this one. The fault injected — returning `None` for both creating
columns, which is the pre-fix code — reports three nameless columns where one
is allowed.

### a_records_coordinate_is_text_however_it_is_spelled

improvise-tom1. `project::cell_display`'s records arm called `records_display`
— which returns a `String` — and then inferred numeric-ness with
`text.parse::<f64>()`. A records coordinate column holds an item **name**, so a
`Year` category with items `2024`/`2025` projected a coordinate the browser
right-aligns as a number.

Nothing in the suite could see it: every records fixture in the tree names its
categories `Region`/`East`/`North`, so the guess happened to be right. The
fixture here is the whole test — a category whose items are spelled like
numbers — and the measure column beside it is asserted numeric in the same
test, so a producer that answered `false` for everything would not pass.

Fault injected: the `parse::<f64>()` guess restored verbatim. It reports the
coordinate as numeric.

### a_staged_drill_edit_reaches_its_own_session_and_no_projection

The user's ruling, pinned in both directions (improvise-tom1). A remote client
must NOT see another client's staged drill edits, so `grid_of` must project the
committed value; the session that staged one must still see its own typing,
through `App::display_value_of`.

Asserting only the second half would pass against a producer that leaked, and
asserting only the first would pass against one where staging did nothing at
all — so both are here. The fault injected is the leak: teaching
`cell_display`'s records arm to prefer `drill_state.pending_edits`, which is
the plausible "fix" this ticket exists to forbid. The projected cell then reads
`999` instead of `1`.

The fixture takes its `DrillState::records` from the layout the app really
built, because an empty frozen snapshot silently empties the drill grid and the
test then asserts nothing about a cell that does not exist — the first version
did exactly that and panicked on `expect`.

### a_page_axis_carries_its_items_and_a_filter_axis_does_not

The producer half of improvise-p1ym, added during self-review because
`a_filter_axis_is_projected_but_not_cycled` builds its own `PageAxis` by hand —
a test that fabricates the very data the code under test is supposed to
produce, which is this repo's characteristic defect. Only this one can say what
`grid_of` actually emits.

The fixture puts `Quarter` on `Axis::Page` and `Year` on `Axis::Filter`, which
is the shape a drill makes (gotcha 27), and asserts the whole projected list in
one comparison so a producer that emitted items for **everything** or for
**nothing** both fail. Fault injected: `let navigable = true`, which is
`filtering_cats()` leaking into the protocol — the exact trap the ticket names.
`Year` comes back carrying `["2024"]`.

### a_records_grid_has_no_column_for_a_formula_target

A records grid has no column for a formula target that is **only** a formula
target, because `GridLayout::records_columns` builds its measure columns from
`Category::ordered_item_names` — the explicit `_Measure` items — while
`add_formula` adds none (gotcha 7). Fault injected: swapping
`ordered_item_names` for `measure_item_names` in `records_columns`, which is a
plausible tidy-up — a `Profit` column appears.

**This test's original premise was FALSE, and the review caught it.** It was
written claiming the records branch of `cell_id` was *safe* because "a formula
target lives only in `measure_item_names`". An item can be **both** an explicit
`_Measure` item and a formula target, and improvise-poia — in the same PR —
made the `+` column that builds exactly that reachable. So this test asserted
the accident in the one configuration where the accident holds, and its own
failure message named a repair for a case it could not reach. The real guards
are `a_records_column_a_formula_owns_is_derived_not_a_record` and
`typing_into_a_records_formula_column_is_refused_on_the_key_path`; this one now
claims only what it can see.

### a_records_column_a_formula_owns_is_derived_not_a_record

The hole improvise-ir7l's first fix left, found by adversarial review. A measure
can be **both** an explicit `_Measure` item and a formula target — the records
`+` column creates precisely that — and `cell_id`'s records branch asked
`formula_for` about nothing, so it projected `CellId::Record{measure:"Profit"}`.
A write there is accepted, stored, and then shadowed forever by `formula_cache`,
which is the defect the ticket exists to close, surviving in records mode in
**both** front ends since `editable` is `Record`-only.

The repair asks the document about the record's **real** coordinate —
`GridLayout::record_cell_key`, the one mapping from a records position to a
model cell — never the synthetic `(_Index, _Dim)` key, which names no measure
and so can never match a formula. Fault injected: deleting the lookup, which is
the pre-review code. `Profit` reads back as a `Record`. The `Revenue` column is
asserted still writable in the same test, so a producer answering `Derived` for
everything could not pass.

### typing_into_a_records_formula_column_is_refused_on_the_key_path

The same defect through `App::handle_key` and no test API at all: `R`, walk to
`+`, type `Profit`, Enter, Esc, then `i` on the column just created. It exists
because the projection half and the `EditOrDrill` half are **separately**
breakable, and a unit test of either would have missed the other — proved by
injecting each alone. Restoring `cell_id` while leaving `CmdContext::
owning_formula` reading the raw key fails on the mode assertion; the reverse
fails on the identity assertion.

One fixture note that cost a debugging round: `CommitAndAdvance` **re-enters
editing** on the cell it advances to, so the mode after Enter is
`RecordsEditing` and an `i` there is a literal character. Without the Esc the
test passes its mode assertion for the wrong reason — nothing happened at all.

### a_save_tells_the_client_the_document_is_clean

improvise-z754's own stated failure mode, shipped: "every client showing a `[+]`
for a document that has been saved". `Host::save` applied `effect::Save` and
returned only a possible `Rejected`, never projecting, so the dirty marker
cleared **outside** any projection and no client was ever told.

The z754 test that was supposed to cover this called `sessions.project_all` by
hand — which is exactly the call the production path omits. It is not a test
that compares a thing with itself; it is a test that exercises a call the
deployment never makes, which is the subtler shape. This one goes through
`Host::handle(ClientFrame::Save)`.

Two assertions guard the fixture: the edit must first report `dirty: true`, and
`app.model_state.dirty` must really be `false` afterwards — otherwise a host
that saved nothing would pass by projecting nothing. Fault injected: returning
an empty vec from the success arm. No `Document` update is emitted.

### the_automatic_persist_tells_the_client_the_document_is_clean

The worker's half of the same defect, and a different cause: `on_frame` built
`out` via `frames_for` and *then* called `persist()`, so the projections were
built while `dirty` was still true. The browser is the deployment where this
bites hardest, because nothing there ever sends `ClientFrame::Save` — the worker
persists on its own behalf (gotcha 45), so this was the only path.

Fault injected: dropping the re-projection after `persist()`. The client's
`DocumentCache` stays dirty. `MainThread` grew a `document` mirror and a
projection dispatch for it, rather than a bespoke reach into the worker — the
point is what a client folding the real stream ends up holding.

**A first version asserted only `!document.dirty`, and `DocumentCache::default()`
is already clean** — so it could not tell "told it cleared" from "never told
anything", and passed under an injection that stopped projecting the widget
entirely. Found in re-review. It now records **every** marker the client is
told and asserts `[true, false]` for the edit, with the bootstrap's own clean
marker cleared first so the sequence is about the edit rather than the
handshake. The stronger form fails under both injections.

### a_refused_or_duplicate_add_category_moves_no_counter

improvise-75ld. `App::apply_effects` bumped `layout_version` and raised
`model_state.dirty` off the `EffectKind` **before** the `abort_effects` check, so
the refusing effect counted itself. Two arms, and the second is the reachable one:
a 13th category is refused (`MAX_CATEGORIES`) and aborts, while a *duplicate*
`add-cat Region` returns `Ok` and does not abort at all — a pure no-op that still
refused every other client's next `Command::Model` with `Rejection::LayoutMoved`
and made the browser worker rewrite the whole OPFS document (gotcha 41).

Injected fault: put the two bumps back on the declared kind, inside the loop
(`if matches!(kind, EffectKind::Model(ModelChange::LayoutGeometry))`). Both no-op
arms fail; so do `neither_shape_counter_moves_for_an_accepted_no_op`,
`both_shape_counters_move_for_every_reshaping_command` and
`a_configuration_change_that_draws_nothing_is_still_a_document_change`, with 450
other tests still passing — the build ran.

The positive arm is not decoration: without it, "never bump anything" passes the
two no-op arms. It also fails under a second injection — dropping the
`document_config` comparison — because `AddCategory` is `LayoutGeometry` and the
declared half of `dirty` no longer covers it.

### both_shape_counters_move_for_every_reshaping_command

improvise-pdvm. `AddItem`, `AddItemInGroup` and `RemoveItem` declare
`ModelChange::CellData` and reshape the grid. Since improvise-9mo9 they call
`App::reset_cursor`, so `cursor_epoch` moved while `layout_version` did not: a
client was told its cursor was invalid while its layout basis tier still read
current, and a batch composed against the old shape was still accepted.

It drives `reshaping_commands()` — the same table
`every_reshaping_model_command_invalidates_every_cursor` drives, extracted into a
shared function precisely so the two cannot answer for different cases — and
asserts the grid shape really moved before asserting either counter, so a case
that reshapes nothing fails loudly rather than proving nothing.

Injected fault: as above. Fails naming `SetAxis`, then `AddItem` once the earlier
cases are removed.

### a_configuration_change_that_draws_nothing_is_still_a_document_change

The third case the shape/config split creates, and the one that forced two
fingerprints rather than one. `hide_item` records a name whether or not the model
has one, and `SetNumberFormat` writes the file while moving no row or column: both
are `dirty` and neither is `layout_version`. A single fingerprint covering the
persisted configuration made a ghost hide bump the basis; one covering only the
drawn shape let a format change go unsaved.

Injected fault: remove the `document_config` comparison in `observe_reshape`.
Fails naming `HideItem { category: "Region", item: "NoSuchItem" }`.

### a_reshape_of_one_view_leaves_a_client_on_another_view_alone

improvise-nynk. improvise-jytm made "which view a client is on" per-client;
`Sessions::adopt_cursor_epoch` was not updated and still reset **every** mirror
and returned one `Broadcast::ToAll(Command::ResetCursor)`. A reshape of view
`Default` therefore threw away the cursor of a client sitting on `Second`, which
nothing reshaped.

The fixture is two views of one document with different axes, and the bystander
switches to `Second` and scrolls there before the asker reshapes `Default`. The
positive arm is kept in the same test — the asker is on the reshaped view, is
reset, and is told — because without it "never invalidate anyone" passes.

Injected fault: drop the `cursor_invalidated` filter (`.filter(|_| true)`). Fails
naming the bystander's thrown-away cursor, with 453 other tests passing.

### an_unknown_reshape_set_invalidates_everybody

The fail-safe arm of the same narrowing, tested separately because it is the half
gotcha 24 calls strictly worse than doing nothing. An empty reshaped set means the
batch moved a cursor with no view observed to change shape; the answer must be
"everybody", not "nobody", or the diff believes each client already holds
origin-positioned content and sends nothing, leaving them stale permanently.

Injected fault: delete the `reshaped.is_empty() ||` disjunct. Fails by name. That
injection is why the predicate is a named function rather than a closure — a
fail-safe nothing can point at is a fail-safe that gets tidied away.

### a_batch_that_changes_the_layout_and_then_saves_ends_clean

A regression introduced by the first version of the observed-`dirty` change and
caught in review of PR #70, not by the suite — the workspace was green at 1400 with
this broken. `observe_reshape` ran **after** the effect loop and re-raised
`model_state.dirty` from the `document_config` comparison, while `Save` had cleared
it **inside** the loop. So gotcha 45's stated ordering property — "`Save` clears it
in loop order, so a batch that edits and then saves ends clean" — survived for
`CellData`/`WholeWorkbook` and was broken for `LayoutGeometry`.

`[SetNumberFormat, Save]` reported a successful save and left the document dirty:
the TUI draws "Saved to …" while `Quit`'s check refuses a later `:q`, and the
browser worker — whose persist trigger *is* the flag — pays a whole-document OPFS
rewrite. Reachable from `:format ,.2f . w` and from any `Binding::Sequence` pairing
a geometry effect with a save.

The fix is position, not logic: the comparison moved inside the loop, taken after
each `LayoutGeometry` effect against the value held after the previous one. Both
arms are in the test — `[SetCell, Save]` never left the loop, so the geometry arm
alone could not say whether the rule is general — and each asserts the save really
happened (`last_batch_aborted`, and a status line starting "Saved") before
asserting the flag, or a failed save would pass the dirty assertion for the wrong
reason.

Injected fault: move the comparison back after the loop. Fails naming the layout
arm, with 454 others passing.

The lesson worth carrying: **an observation that is correct about *what* changed
can still be wrong about *when* it is read.** Everything above this entry argues
that observing beats declaring; this is the cost of that move, and it lands
precisely on the one effect that *unsets* what the observation sets.

### the_published_basis_names_the_records_the_grid_draws_and_no_others

improvise-23qd. The set of records a `Projection::Basis` carries was derived
**twice** — `ClientSession::basis_update`, the producer, and `server_truth`, a
`#[cfg(test)]` helper that is the **yardstick every fold test measures against**.
So a change to which records get published moved the thing and its own measure
together, and every test stayed green whatever the change was. Not a helper that
fabricates an input; a helper that fabricates the expected *output*.

Both now call `published_basis`. Note what that alone does **not** buy: the
ticket's stated proof ("alter the filter and watch the basis tests fail") is
impossible once one function feeds both sides — they still move together. A shared
derivation removes the drift and removes the yardstick with it.

So the yardstick is spelled out rather than derived: this test names North, South
and East by hand for `app_with_a_reshapeable_grid`'s three records over six cells,
and asserts that a grid where every cell is an `Aggregate` publishes **none** —
because an aggregate names no record, so there is nothing to claim a version for.

Fault injected: `.take(2)` on the one surviving filter — the ticket's exact "change
which records get published" — fails this test **and nothing else in the crate**.
A second injection, `record_version_of(..) + 1`, is caught instead by the two
end-to-end `basis_check` tests, which is the split worth knowing: the counters have
other guards, the membership had none.

The fixture carries its own premise check. The first version assumed a new category
lands on `Axis::None`; it lands on `Column`, and the assertion said so instead of
the test quietly proving nothing.

### an_aborted_batch_bumps_layout_version_for_what_it_did_run

improvise-2avl, with its negative twin `an_aborted_batch_bumps_nothing_for_effects_it_never_ran`.
PR #70 moved the `layout_version` bump from "once per `LayoutGeometry` effect,
inside the loop" to "once per batch, after the loop". Gotcha 40 records the in-loop
position as load-bearing — an aborted batch must not claim effects it never ran —
and the author's argument for why end-of-batch is safe *for `layout_version`
specifically* (nothing ever clears it, so an end comparison sees only what ran,
whereas `dirty` must stay in the loop because `Save` unsets it mid-batch) is sound
and was written into `context/code/command.md`. **An argument is not a test**, and
neither arm had one.

Both arms are proved **independently**, which is the part that earns them: bumping
unconditionally fails only the second test, and skipping `observe_reshape` on abort
fails only the first. A single injection failing both would not have shown that the
two tests see different things.

Each fixture asserts `app.last_batch_aborted()` before asserting anything about the
counter, because a batch that did not actually abort would pass both arms while
testing neither.

### sorting_the_store_in_records_mode_invalidates_the_cursor

improvise-ni6k names `SortData` as its **sharpest** case: it changes neither the
row count nor any label, and every position still names a different record. A guard
written as "did the shape change" misses it; `pending_edits` are keyed by row index
(gotcha 1), so row 0 being West and then East is gotcha 18's misdirected-write
disease arriving through a different door.

`view_frame` therefore hashes each record's **coordinates** alongside the axis
entries, because records-mode `row_items` are positional — `0,1,2…` whatever order
the store is in. Fault injected: deleting those three lines fails this test and
**nothing else in the workspace**.

**The fixture trap, which is the durable part.** PR #78 first shipped without the
coords hash, on the stated ground that no fixture could be built where `SortData`
reorders anything. That was wrong, and the reason is worth more than the fix:
`ToggleRecordsMode` pushes `effect::SortData` as its **first** effect, so a
records-mode fixture built by pressing `R` has an already-sorted store *by
construction* and cannot demonstrate anything order-sensitive. The store itself was
fine all along. So this fixture sets `_Index`/`_Dim` directly rather than going
through the key path.

The test asserts the count and the labels are **unchanged** before asserting the
cursor moved, so it cannot pass by accidentally becoming an ordinary shape change.

### paging_with_nothing_pruned_invalidates_no_cursor

The negative twin of `paging_across_a_prune_that_drops_a_row_invalidates_the_cursor`,
and the reason `view_frame` hashes **no page selection**.

improvise-7kvw is real only under `prune_empty`: a page flip changes which records
are visible, not which categories frame them, so with pruning off a grid position
keeps naming the same cell. PR #76 ruled that homing the cursor on every `[`/`]`
regardless — `[`/`]` being navigation a user does constantly — is **worse than the
hole**. Hashing the page selection does exactly that.

The seductive argument for hashing it anyway is that two pages with identical row
labels really are different data. True, and irrelevant: that is `record_versions`'
question, and `view_frame` answers the cursor's. A basis argument applied to a
cursor decision is how this regression got written in the first place.

Fault injected: restoring the three `page_selections` lines fails this test and
nothing else. The positive arm keeps passing either way, because
`with_frozen_records` applies `prune_empty` and so already sees a dropped row in
the axis entries — which is why the negative arm is the one that had to exist.

### a_write_that_prunes_a_row_out_of_another_clients_view_invalidates_it

improvise-mcq4's sting, found by the adversarial review of PR #78. The mcq4 fix
narrowed the frame fingerprint from every view to the active one, on the strength
of a measurement showing the narrowing cost nothing. **It bought its speed partly
with correctness**: a `CellData` write that prunes a row out of *another* client's
view became invisible. `view_shapes` is config-only — it hashes the `prune_empty`
flag, not which rows survive — so `last_reshaped_views` is empty, the one-view
frame covers only the host's, and a client on the other view is never sent
`ResetCursor` while its grid loses a row underneath it. That is gotcha 24's
"strictly worse than doing nothing": the diff then believes the bystander is
current, so it stays stale **permanently**.

The repair keeps the narrow one-view frame on the `LayoutGeometry` arm — the
held-key path improvise-748f exists for, where other views cannot reframe because
`view_mut()` writes only the active one — and restores the all-views frame on the
`CellData` arm. Measured: **0%** on the held-key path, **+5.7%** on an 18k-cell
two-view `SetCell`, scaling linearly in view count.

**Writing this test found a second bug, which was masking the first.** The
precondition failed at first: `grid_in` built a non-active view with
`GridLayout::new`, which does **not** apply `prune_empty`, so a client on another
view was projected three rows where its own config gives two. Their grid never
pruned, so it never lost a row, so the staleness was unobservable — a bug hiding a
bug. It is the third instance of the `new`-versus-`with_frozen_records` trap
(`view_frame` was the second), and fixing it broke no other test.

Fault injected: removing the all-views arm fails this test, plus
`sorting_the_store_in_records_mode_invalidates_the_cursor` (`SortData` is
`CellData`) and both cursor-audience tests — four in all.

### a_switch_then_a_geometry_effect_compares_within_one_view

Found by the same review. `frame_basis` was taken once at batch open against
`view_state.active_view`; `SwitchView` is `ViewOnly`, so it did not re-base, and
the next model effect hashed the **new** view and compared it against the **old**
view's hash. A near-certain mismatch, so a spurious `reset_cursor()` and a wrong
name in `last_reframed_views`.

Today's instances are benign — `DrillIntoCell` and `EnterRecordsMode` both emit
`[CreateView, SwitchView, SetAxis, …]` in one batch, but against a view they just
created. The harmful case needs a switch to an *existing* view with a saved cursor
plus a model effect in the same batch, which a `.`-chained `:` line reaches. Low
severity, unsound mechanism.

The basis is now keyed by view name and re-bases rather than comparing when the
name moves. Fault injected: dropping the `was == self.view_name()` guard fails
this test, with `cursor_epoch` moving for a batch that reframed nothing.

improvise-9c9o moved *where* the re-base happens, and this test was unchanged by
it because it names the property, not the mechanism — the re-base now fires on
the switch rather than on the next geometry effect, and a batch that reframes
nothing still invalidates nothing.

Its fixture carried a line that did nothing at all:
`views["Second"].clone_into(&mut views["Default"].clone())` clones into a
temporary that is immediately dropped. It read as the thing making the two views
differ, so had `create_view` ever started copying the active view's config this
test would have gone vacuous with that line standing there looking like it
prevented exactly that. It is deleted and replaced by the assertion it was
pretending to be: the two views' `view_frame` hashes really differ.

### a_geometry_effect_after_a_mid_batch_switch_still_invalidates_the_cursor

The other side of the guard above, and introduced by it. PR #78 round 1 compared
frame hashes across a view change, which threw a cursor away spuriously; round 2
added `was == self.view_name()`, which fixed that and also **exempted the first
geometry effect after a mid-batch `SwitchView` from the check altogether** —
there is no basis for the new view yet, so the code re-based and skipped.
Passes on pre-#78 `main` and on round-1 head; fails at the merged head.

Direction is why this is worth more than the bug it replaced. Round 1 kept a
cursor it should have thrown away; round 2 threw away nothing it should have
kept — which is gotcha 35's expensive side, since a stale cursor has nothing in
the stream to correct it where a spurious invalidation costs one projection.

The fix is the ticket's: **the basis follows `active_view`**, re-based in the
loop as soon as the name moves, rather than at the next geometry effect. It did
**not** need `SwitchView` to stop being `ViewOnly` and it is not a protocol
change — the basis is `App`'s own bookkeeping, and nothing about it is on the
wire. The `was == self.view_name()` guard is gone from the comparison because it
is now unreachable: by the time the geometry arm runs, the name and the basis
agree by construction.

What it changes elsewhere, and why that is fine: `DrillIntoCell` and
`EnterRecordsMode` both emit `[CreateView, SwitchView, SetAxis, SetAxis, …]`, so
they now reset the cursor on the *first* `SetAxis` where they used to skip it and
reset on the second. Same outcome, one effect earlier. `DeleteView` is the one
`LayoutGeometry` effect that moves `active_view` itself; the re-base fires for it
too, so the geometry arm compares the new view against itself and resets nothing
— which is what the old skip did.

The fixture measures the row count the transpose must move rather than spelling
it: a first version asserted `row_count() == 1` and the real answer is 2, which
is the shape of a precondition that would have gone quietly wrong later. It
fails on the unfixed tree at the epoch assertion, `left: 0 right: 0`.

### a_frozen_drill_does_not_take_its_opening_frame_from_the_held_layout

improvise-mcq4's fix reads the batch's opening frame off `App::layout` rather than
building a second pruned layout — that is where the 4.1×→1.9× came from. It is
guarded on `drill_state.is_none()`, because a drill freezes the record list and the
held layout is then **not** what `view_frame` would build for that view.

**The guard was correct, load-bearing, and tested by nothing**: reverting it to the
unconditional form left all 1439 tests passing. Textbook
`doctrine(a_careful_comment_is_weak_evidence_that_anything_checks_it)`, and it was
the most carefully argued comment in the PR.

The fixture freezes one record where the live query finds three — asserted, so a
freeze that did not narrow could not pass — then runs a `SetNumberFormat` batch,
which reframes nothing. Without the guard the opening basis is the frozen layout
and the in-loop hash is the live one, so they differ and the cursor is reset for a
batch that changed no row. Fault injected: replacing `self.drill_state.is_none()`
with `true` fails this test and nothing else in the workspace.

### a_clients_grid_does_not_depend_on_which_view_its_host_is_on

improvise-t6db. The one-line `grid_in` fix landed in PR #78 with no test of its
own; this is the general property the reviewer specced — **host-view
independence**: `grid_in(app_on_A, "B", vp) == grid_in(app_on_B, "B", vp)` for
every ordered pair of views. It needs no second derivation of what a view's
config means, which is why it is a property and not a second copy of
`GridLayout`.

**The trap is the fixture, not the assertion.** The reviewer's first version made
the two views differ by turning `prune_empty` **off** in the second — plenty of
differing config, and vacuous: with nothing to prune, `GridLayout::new` and
`GridLayout::with_frozen_records` agree and the property passed on the buggy
tree. So `app_with_three_views_and_an_empty_row` gives Service no cell at all and
leaves all three views pruning, and the test asserts every view really is two rows
before comparing anything — a fixture that stopped pruning would fail there rather
than passing quietly.

The property is stated with `drill_state.is_none()` asserted, because a frozen
drill snapshot is the one thing `grid_in` cannot reproduce for a view the host is
not on (`with_frozen_records(.., None)` on the elsewhere path).

Fault injected 2026-09-21: restoring `GridLayout::new` in `grid_in` — the pre-#78
line — fails this test at the `assert_eq!`, not at the precondition. It also fails
`a_write_that_prunes_a_row_out_of_another_clients_view_invalidates_it` and
`a_hosts_none_cats_leak_into_another_views_formula_columns`: **three** tests, not
the one an earlier draft of this entry claimed. The #78 fix has more coverage than
it was credited with.

**The `none_cats` asymmetry is REAL, and the first draft of this entry reported it
as unobserved from a fixture that lacked the condition.** That correction is worth
more than the test. The claim was that the fixture put `Product` on None in
`Default` and on Column in `Second`; it did not — `add_category` defaults a new
category to Column and nothing moved it, so all three views had `none_cats`
`{_Index, _Dim}`, the stated input to the hazard was absent, and `Default` and
`Second` were indistinguishable to `grid_in` apart from a `number_format` that is
itself a no-op. An unprobed path reported as a negative result, which is
`doctrine(an_injection_that_does_not_fail_names_an_untested_path)` applied to a
**setup**, and the expensive kind: it stops the next person looking.

Setting `Product` to `Axis::None` explicitly made the property fail on the first
run, and not on the value of a cell — `Second` lost **both `Total` columns
entirely** when the host stood on `Default`. Filed as improvise-ac2z, pinned by
the characterization test below, and the property now skips pairs whose
`none_cats` differ. That skip is a **bug being worked around**, not a carve-out;
it asserts `compared > 0` so a future fixture cannot skip everything and pass.


### a_window_narrows_the_cells_and_neither_the_totals_nor_the_widths

improvise-jmjz. `project::grid_of` evaluated the whole grid roughly three times
per frame — `compute_col_widths` walked every cell for its text and then every
cell again for the column totals it must size, and `grid_of` then evaluated the
window a third time. `ui/viewport.rs::grid_body` now evaluates each cell once and
derives the widths, the totals and the window from that value.

That fusion creates a hazard the three separate passes did not have: one walk now
serves three consumers with different scopes. **Two** faults follow, not one — a
window that leaks into the totals makes a column total the sum of the visible
rows, and a window that leaks into the widths sizes a column by what happens to
be on screen. Neither is visible from the output: the numbers simply get smaller,
consistently.

**The fixture took three attempts, and what each one was blind to is the entry.**

*First*: rows of `1`, `20`, `300000` under a measure named `Revenue`, asserting
`col_widths[0] > "300021".len()`. It could not see a windowed **width** at all —
`Revenue` is 7 characters against a 6-character value, so the assertion passed on
the **header** alone. Worse, its other half was structurally incapable: comparing
`grid_body`'s widths against `compute_col_widths` cannot test windowing, because
`compute_col_widths` **is** `grid_body` with an empty window. Windowing degrades
both sides identically, both collapsing to the `MIN_COL_WIDTH` clamp. **The same
function twice is not a check, and it reads exactly like one.** Measured by the
PR #92 review: `if drawn { widen(..) }` left **1476 passing, 0 failing**.

*Second*, the reviewer's: `1`, `888888888`, `-888888888` with a one-character
measure name, so the total **cancels to 1.0** and neither it nor the header can
supply the width. This fixed the width half and broke the other: the windowed row
is `1.0` and the cancelled total is *also* `1.0`, so the **totals** assertion went
blind to the row-window fault, which the widths assertion then happened to catch.
Two assertions, one no longer discriminating, and nothing saying which.

*Third*, and current: the same values with the **large ones first**, so the window
holds `888888888` while the total is `1.0`. Each injection now fails on the
assertion that names it:

- `for ri in rows.clone()` — the whole walk windowed → **totals**
  `[888888888.0]` against `[1.0]`
- `if drawn { widen(..) }` — widths only → **widths** `[10]` against `[11]`

`10` rather than `5` is a smaller gap than the reviewer's fixture gave and a
better witness: the missing character is the **minus sign**, which only
`-888888888` carries and the window does not contain. The test also asserts the
window really is narrower than the grid, so a fixture that stopped windowing
would fail at the precondition.

**What this test does not do is prove the speedup**, and no test can — the outputs
are identical by construction, which is why the whole suite passed unchanged
across the change. The evidence is the benchmark, before and after, in
`context/code/command.md`: about 3x, the number of passes removed.

### a_formulas_columns_survive_a_host_that_aggregates_differently

improvise-ac2z, **fixed**. This replaces the characterization test
`a_hosts_none_cats_leak_into_another_views_formula_columns`, whose message said
that failing meant the bug was fixed; it did, and it is gone.

`App::rebuild_layout` calls `recompute_formulas(&none_cats)` for the **host's**
view, populating `Model::formula_cache`. `project::grid_in` on the elsewhere
path built the other view's `GridLayout` against that same cache, which is keyed
by the coordinates the recompute used: with `Product` on None in the host's view
the `Total` cells are keyed *without* a `Product` coordinate, a view with
`Product` on Column asks for `(Region, Product, _Measure=Total)`, finds nothing,
and — `prune_empty` being on — drops both `Total` columns. Not a wrong number:
absent columns.

  host on `Second`, projecting `Second`   6 columns, `Total` present
  host on `Default`, projecting `Second`  4 columns, no `Total` at all

The fix is `grid_in::rederived_for`: where the target view aggregates over
different categories than the host's, it clones the model and recomputes the
formulas for **that** view's `none_cats`. Three conditions gate the clone —
another view, different `none_cats`, and at least one formula — so the common
case pays nothing and a formula-free document pays nothing ever.

**This is an interim, and the cost is the reason.** A clone of the whole model
plus a fixed-point recompute, per projection, on exactly the path that is broken
today. The ticket's first candidate — keying `formula_cache` by `none_cats` so
one `Model` can hold more than one view's derived values — is the durable answer
and is filed separately; it is an `improvise-core` change and did not belong in
this wave's ground. What is bought meanwhile is correctness on a path that was
silently wrong, `doctrine(the_elsewhere_path_is_only_as_correct_as_the_inputs_it_can_rebuild)`.

**Fault injected**: `rederived_for`'s condition forced to `false`. Two tests
fail — this one, and
`a_clients_grid_does_not_depend_on_which_view_its_host_is_on`, whose `none_cats`
carve-out was removed at the same time and which is the general property this
one is an instance of. The test asserts its two **preconditions** first — that
`Second` on its own really draws the formula's columns, and that the two views
really do aggregate over different categories — so a fixture that lost either
fails rather than passing on two empties.

### a_view_named_command_lands_on_the_view_it_names

improvise-lltt, the oldest open P1 here, opened by improvise-jytm making
`ViewState::active_view` authoritative per client.

Nine `ModelCommand` variants configure a view — `SetAxis`, `CycleAxis`,
`TransposeAxes`, `SetPageSelection`, `HideItem`, `ShowItem`, `ToggleGroup`,
`TogglePruneEmpty`, `SetNumberFormat` — and each named its view *implicitly*:
`model_effects` built an effect that wrote `App::view_mut()`, which is the
**host's** slice. A client that had switched to `Second` therefore moved
`Default`: its own grid did not change, and every other client's did.

The fix is the protocol change the ticket asks for. Each of the nine carries a
`view: String`, `model_effects` maps it to `ViewTarget::Named`, and the nine
effects carry a `ViewTarget` rather than reading an ambient one. There is no
operandless form left: `App::view_mut()` is **deleted**, replaced by
`view_mut_of(&ViewTarget)`, so a future geometry effect has to say which view it
writes. The host's own key path passes `ViewTarget::Active`, which is right
because the TUI *is* the host and has no session to ask.

**Fault injected**: `model_effects`' `SetAxis` arm rebuilt with
`ViewTarget::Active` (the pre-fix behaviour, with `view` bound and discarded so
it still compiles). The test fails on `the asker's own view never moved` —
`Second`'s `Channel` stayed on `Row` while `Default`'s moved to `Page`.

The fixture must seat the asker somewhere the host is not, and the two views must
disagree about the category being moved: `Channel` is `Axis::None` in `Default`
and `Axis::Row` in `Second`, and the command asks for `Axis::Page`, so both
assertions separate the fixed code from the broken code in both directions.

### a_reframe_of_a_view_the_host_is_not_on_invalidates_its_clients

improvise-lltt's second half, and the part the ticket did not predict. Naming the
view made a reframe of a view the host is not on reachable for the first time,
and `Sessions::adopt_cursor_epoch` could not see one: it returned early unless
`App::cursor_epoch` had moved, and the epoch is bumped only by `reset_cursor`,
which `apply_effects` calls only for the **host's own** view. So the clients on
the reframed view were reset by nobody and told nothing — gotcha 24's
strictly-worse-than-nothing state, permanently stale rather than stale until the
next scroll.

`App::last_reframed_elsewhere` is the second trigger. Its opening frames are
taken lazily inside `App::view_mut_of`, which is the one door such a write goes
through, so the held-key path still hashes exactly one view: the wide-observation
cost improvise-mcq4 declined is not paid.

**Fault injected**: the gate restored to `if !host_moved { return Vec::new(); }`.
The test fails on `a client on the reframed view was never told`.

Two negative arms, and they are not decoration. A third client left on `Default`
must keep its offset, or "invalidate everyone" passes; and
`paging_with_nothing_pruned_invalidates_no_cursor` is what forbids relaying
against `last_reframed_views`, which is a union with the *reshaped* set and so
fires for a page selection that pruned nothing. That test failed during
development and is why the elsewhere set is precise.

### a_command_naming_a_view_the_document_has_not_got_writes_nothing

`App::resolve_view` falls back to the document's default for a name the document
has not got, which is right for *reading* — a slice naming a deleted view must
not panic in `rebuild_layout` — and wrong for **writing**, where it would move a
view nobody asked about. `view_mut_of` therefore looks the name up directly and
returns `None`.

**Fault injected**: `view_mut_of`'s `Named` arm routed through
`self.resolve_view(name).0`. The test fails on `a name the document has not got
fell back onto a view that has one`.

The window this closes is narrow — `Sessions::adopt_deleted_views` already moves
a mirror off a deleted view and tells it — but it is a race, not an
impossibility, and the silent alternative is a write to the wrong document state.
Not closed here: the client hears nothing. A `Rejection` variant for it is filed
rather than smuggled in, since `Rejection` is wire vocabulary.

### the_diff_tells_a_client_the_grid_became_a_records_grid

The producer half of `GridCache::records` (improvise-e5pf). `grid_of` setting
the field is exercised by `leaving_a_panel_in_records_mode_lands_where_the_host_lands`;
the **diff** is a separate arm, and a client that connects before records mode
is entered learns only through it.

Both preconditions are asserted — the fixture starts on an ordinary grid, and
`R` really reaches a records one — so the comparison is between two answers
that had something to disagree about.

**Fault injected**: the `records` arm deleted from `grid_updates`. The test
fails on `the client is never told`. The negative arm (an unchanged flag is not
resent) is what stops "always push it" from passing.

### a_batch_that_leaves_records_settles_out_of_records_normal

improvise-4dpj, gap 1. `settle_normal_mode` matches `Normal | RecordsNormal`,
and replacing that with a one-way `AppMode::Normal` left the **entire workspace
suite passing**. The existing crossing test,
`a_batch_that_leaves_records_and_then_sets_normal_lands_in_normal`, drives the
layout both ways, but both of its branches end in `enter-mode normal`, so the
mode handed to the settle is `AppMode::Normal` in both directions. It
discriminates "settles at all" and "settles against the ending layout", and not
"settles from **either** mode".

This drives the production repro: from a records grid,
`apply_effects([SetAxis { "_Index", Axis::None }])` leaves
`layout.is_records_mode()` false with the mode still `RecordsNormal`, so the
`RecordsNormal` arm is live and not defensive. Both the records-ness of the
grid and the mode before the batch are asserted as preconditions, so a fixture
that stopped entering records mode fails loudly rather than proving nothing.

Injected fault: `if matches!(self.view_state.mode, AppMode::Normal)`. Fails by
name, `left: RecordsNormal, right: Normal`. It is the only test in the
workspace that fails.

### switching_views_across_records_ness_settles_against_the_view_switched_to

improvise-4dpj, gap 2 — raised by ChatGPT Codex against PR #91 and verified
correct on `main` before being tested. `SwitchViewAtCursor` emits
`[SwitchView(target), ChangeMode(Normal)]`, and this is a **different path**
from the one gap 1's siblings drive: that test toggles records mode explicitly,
where a view switch crosses records-ness only because it is auto-detected from
the axes (gotcha 1). Nothing in the batch says "records".

Driven in both directions, with the crossing asserted before and after so a
fixture that stopped crossing fails rather than passes. Injected fault: move
`settle_normal_mode()` back above `rebuild_layout()` at the end of
`apply_effects` — the exact regression PR #91's round 2 fixed. This test fails
by name; so does gap 1's, along with six pre-existing records-mode tests
(measured in review; an earlier draft of this entry said five).

**Scope note, stated because an overclaim is worse than a gap**: this drives
the *effects* `SwitchViewAtCursor` emits, not the view-panel key path that
composes them. Reaching the panel needs a `view_panel_cursor` on a `CmdContext`
built in `command/cmd/panel.rs`, which is another wave's ground this round. The
command's own coverage lives there; what is unexercised is the two joined.

### the_layout_geometry_cases_are_every_wire_effect_of_that_kind

improvise-26dm. `layout_geometry_effects_actually_change_document_config` was
guarded by `assert_eq!(cases.len(), 13)`, which catches a case being *deleted*
from the list and never one being *absent* from it — and the two are written
identically as an `assert_eq` over a number. The adversarial review of PR #78
added a fourteenth `pub struct` declaring `LayoutGeometry`, pointed at by
nothing, and the whole suite stayed green.

improvise-5ubs's mechanism cannot reach it: an effect is a struct implementing
a trait, so serde has no derived enumeration to diff against. `model_effects`
maps each `ModelCommand` variant back to the effect it came from (gotcha 38),
so the list is diffed against the `LayoutGeometry` effects reachable through
`ModelCommand`. The two sets are equal today, which is the evidence the
substitution lost nothing.

**The right-hand side is bounded by the derive, and that took a second round.**
Review of PR #96 injected the deletion of `ModelCommand::TogglePruneEmpty` from
`every_model_command()`: this was the **only** test in the workspace that
failed, and `every_model_command`'s doc comment claiming the list was
compiler-forced was wrong — `variant_name`'s exhaustive match forces an arm
*there*, not an entry in the list. ChatGPT Codex raised the same thing on the
PR. `the_enumeration_names_every_variant_the_derive_declares` closes it, so the
chain is derive → enumeration → `model_effects` → case list with every link
checked. The assertion here still names **both** directions, because a short
right-hand side reads as a spurious case on the left and invites deleting it —
degrading both sides together.

Injected fault: drop `TogglePruneEmpty` from `layout_geometry_cases()`. The old
count said `12 != 13`; this says
`right: [… "TogglePruneEmpty" …]` and **names the effect**.

**Residual, stated rather than closed**: this is total over the wire and not
over the effects. An effect declaring `LayoutGeometry` that no `ModelCommand`
names is still invisible — which is precisely PR #78's probe. Closing that
needs an enumeration of effects the compiler maintains, and none exists without
a proc macro or inventory-style registration.

### the_enumeration_names_every_variant_the_derive_declares

The bound the entry above rests on. Injection is Codex's own, run end to end: a
`ModelCommand::ProbeGeometry` whose effect declares `LayoutGeometry`, with the
`model_effects` and `variant_name` arms the compiler forces, omitted from
**both** lists. Before this test, the whole workspace stayed green. After it,
the failure is staged and each stage names the probe — first
`every_model_command() never constructs ["ProbeGeometry"]`, then, once the
enumeration is repaired and only the case list is short,
`the_layout_geometry_cases_are_every_wire_effect_of_that_kind` with
`ProbeGeometry` on the right. That staging is the point: the two tests answer
different questions and the injection separates them.

The name reader is shared, not copied: `reduce.rs`'s `effect_type_name` and its
own `the_effect_name_reader_reads_all_three_debug_shapes` are what this leans
on, because a second reader here would be a derivation done twice and could
drift while one copy was repaired. It reads the leading identifier off the
derived `Debug`, `Effect` being a trait object that `type_name` cannot reach.

---

## `command/cmd/panel.rs` — the TUI half of the category-panel defects

Both of these have a twin in `improvise-protocol/src/resolve.rs` whose leaf
name is identical, so both headings are module-qualified (gotcha 48). Neither
defect was browser-specific: the same `Cmd` backs the same key in the TUI, and
confirming that before fixing is what kept the fix out of the client.

### command::cmd::panel::tests::deleting_a_categorys_last_item_leaves_the_cursor_inside_that_category

improvise-uej8 in the TUI. The full argument, including why the rival
"always step back" repair is worse than the bug, is under
`resolve::tests::deleting_a_categorys_last_item_leaves_the_cursor_inside_that_category`
in `clients.md`; it is not repeated here, because two copies of an argument
are two things to keep true.

What this test adds that its twin cannot: the TUI tree carries the three
**virtual** categories ahead of the regular ones (improvise-vjms, gotcha 7),
so its indices are the ones a real user's cursor holds. Row 5 is `Type`'s last
item and row 6 is `Month`'s header; the test asserts both before acting, so it
cannot pass by landing somewhere the escalation was never possible. Inject the
fault by deleting the `effects.extend(retreat_from_deleted_item(...))` line.

### command::cmd::panel::tests::deleting_an_item_with_a_sibling_below_it_leaves_the_cursor_alone

The discriminating input for the rival fix, in the TUI. Fails under an
unconditional step-back, passes under the conditional one.

### command::cmd::panel::tests::deleting_a_categorys_only_item_does_not_arm_the_next_d_with_remove_category

The singleton arm in the TUI. The enumeration and the landing argument are
under the `resolve.rs` twin of this name in `clients.md`; they are not repeated
here, because two copies of an argument are two things to keep true.

What this fixture adds: `Month` is seated **before** `Type` and carries two
items, so the safe landing is found by the **backward** search — the preferred
direction — and lands on a real item rather than falling off the end of the
tree. Its twin exercises the same case with nothing below. Both assert the
consequence (a second `d` composes no `RemoveCategory`) rather than the landing
index, and both fail under the pre-review rule while the two adjacent-case
tests pass.

### command::cmd::panel::tests::cycling_an_axis_from_the_category_panel_says_so_as_the_tile_bar_does

improvise-25bf in the TUI. The category panel and the tile bar are two routes
to one operation and only the tile route told the user, which is
design-principles §5 in miniature — a derivation done twice will diverge, and
the divergence is invisible until someone puts them side by side. The wording
now has one home, `viewmodel::axis_change_status`, called by both routes here
and by both in the browser.

The fixture sets `Type` to `Row` explicitly rather than reading the default,
so the expected `Type → Col` is a literal and not a restatement of the code
under test. It asserts `cat_at_cursor()` is `Type` first, because cursor 3 is
only the first regular category while the three virtual ones lead the tree —
an off-by-one would otherwise pin the wrong category's sentence and still pass.

---

## The view stack is the client's (improvise-gd28)

Two user rulings, 2026-09-21: *"The state drill manipulates (the view stack)
should be local to the client"* and *"Drilling should be an ephemeral view, not
something that gets persisted."* The first one is what these three tests serve.

### a_client_folding_a_switch_records_the_history_the_host_does

**The bug, demonstrated before the fix.** `effect::SwitchView` pushes the view
it is leaving onto `ViewState::view_back_stack` and clears the forward stack;
`reduce_view`'s `Command::SwitchView` arm set `active_view` and **nothing
else**. Both write the same portable field of the same type, and the fold is
supposed to be the mirror's copy of the effect (improvise-cqq/m91) — so a
connected client's own history was empty forever, however many views it visited.

That is the whole reason `view-back` reads as unavailable in a browser: not
that stepping back is hard, but that the client has no stack to pop. The
`declined()` comment attributes it to the drill commit instead, which is a
second, separable reason (see the drill note below).

Written against the unfixed tree it fails with
`left: [] right: [ViewFrame { view_name: "Default", mode: Normal }]`. It drives
the host through `apply_command` and the mirror through `reduce_view`, so it
compares the two writers rather than one writer against itself.

### a_client_stepping_back_lands_where_the_host_lands

The lockstep for the two new commands. `effect::ViewBack` restores
`frame.mode` as well as the view name, so a fixture left in `Normal` cannot
tell "restores the mode" from "never touched the mode" — the mode the default
supplies is exactly the mode a broken restore leaves behind. The fixture
therefore enters `Help`, switches away, drops to `Normal`, and asserts `Help`
comes back. Deleting the `view.mode = frame.mode` line fails it; deleting it
from a `Normal`-only fixture would not.

`stepping_forward_undoes_stepping_back` and
`a_step_off_the_end_of_the_history_moves_nothing` bound the two ends: the
direction is a `Rewind` enum rather than a `bool` so the two arms cannot read
each other's stack, and an empty stack is a no-op rather than a panic.

### a_step_back_cannot_land_on_a_view_the_document_has_lost

The case the client-local stack **creates**, stated where a user feels it: a
frame on my stack can name a view somebody else deleted while I was away.

**This test replaced a guard.** The first version answered it inside
`Sessions::apply_step`, which folded the step and then corrected a dead landing
the way `apply_switch` corrects a refused switch. Two injections did fail it —
but the correction folds a `SwitchView`, which **pushes the dead view straight
back onto the stack**, so pressing `<` again popped it, was corrected again,
and pushed it again: a back button that never progresses. Pruning the history
when the view vanishes (see the entry above) makes the dead frame
unrepresentable instead, which left `apply_step`'s guard with no reachable
input. It was **deleted rather than kept as defence in depth**, because a guard
whose fault cannot be injected is indistinguishable from one that always
passes, and `apply_checked`'s existing fold arm is then the whole behaviour.

The fixture visits `Third` and then `Second`, so the stack is
`[Default, Third, Second]` and only the **top** frame is doomed. Landing on
`Third` is what discriminates: a fixture whose stack held only the dead frame
would land on `Default` and could not tell "skipped it" from "gave up and went
home". Injected fault: drop `forget_vanished_views` from `apply_model` — the
step lands on `Second`, a view the document has not got.

### deleting_a_view_forgets_every_frame_that_names_it

**A defect the client-local stack creates, and a pre-existing one on the host.**
A `ViewFrame` is somewhere you can go *back to*, so a frame naming a view that
has been deleted is an offer the document cannot honour. `effect::DeleteView`
moved `active_view` off the doomed view and left both stacks alone, which was
invisible while no client had a stack at all — the TUI user could already reach
a dead frame with `<`, and making `reduce_view` record history would have given
every browser client the same.

Worse than merely dead: without the pruning, `Sessions::apply_step` corrects
the landing by folding `SwitchView(default)`, which **pushes the dead view back
onto the stack**. Pressing `<` again pops it, is corrected again, and pushes it
again — a back button that never progresses, and one that looks from the
outside like a key that does nothing.

Injected fault: delete the `forget_view` call from `effect::DeleteView`. The
test names the surviving frame. Its fixture asserts the doomed view is on the
stack *before* the delete, so it cannot pass by pruning nothing from nothing.

### a_deleted_view_leaves_the_history_of_a_client_that_did_not_delete_it

The multi-client half, and the one `adopt_deleted_views` cannot cover: that
helper looks at mirrors **sitting on** the deleted view, and a client that
merely *visited* it is sitting somewhere else entirely. The bystander here is
therefore on `Third`, holding `Second` on its back stack — which is also what
makes `ToAll` the right audience and checkable: every client may have been
there, and the asker is not special.

Which views vanished is **observed** by diffing the name set across
`apply_command`, for the reason `adopt_cursor_epoch` watches a counter and
`apply_model` watches `document_generation` (gotchas 24, 39): which effects
drop a view is a per-effect decision, so a table of command variants would be a
second description of it, free to disagree.

Injected fault: drop the `forget_vanished_views` call from `apply_model`. The
bystander keeps its dead frame and the test names it.

### a_refused_switch_puts_no_phantom_on_the_history

**Found by ChatGPT Codex on PR #115, and it is that PR's own regression.** Once
`reduce_view` records history, the *correction* for a refused switch records
history too. Starting on `Default`, the asker folds `Missing` (back gains
`Default`), then folds the correction back to `Default` (back gains
**`Missing`**) — while the mirror, which never folded either, stays empty. So
`<` afterwards lands on a view the document never had, and `forget_view` cannot
help: nothing *vanished*, the name was never real.

**It is reachable**, which is why it was fixed rather than argued away. No
shipped client can invent a view name — `resolve.rs` composes `SwitchView` from
a `ViewSummary` in the projected list — but a client can pick a name that
another client deletes before the frame arrives. That race is exactly the case
`forget_vanished_views` narrows and does not close.

Two things the fix needed, and the test separates them:

- **The mirror folds what the client folded**, refused or not. A correction can
  only undo what both ends hold; guarding the mirror is what let the two drift.
- **The correction is `[ViewBack, ForgetView(asked)]`** — the exact inverse of
  the push, then the refused name off the forward stack it lands on.

Injections, each failing for its own reason:

- **Drop `ForgetView`**: the phantom assertion fires — `Missing` is on the
  history.
- **The rival fix, `SwitchView(leaving)` instead of `ViewBack`** (which is what
  the review's own wording suggests, and what the code did before): no phantom,
  but `left: [Default, Second]` against `[Default]` — the redundant frame the
  correction pushed. That injection is the reason the test asserts the back
  stack is *restored* rather than merely free of the phantom.

**One residue, asserted rather than hidden**: the optimistic fold clears the
forward stack and no correction can refill it, so a refused switch costs the
redo history — exactly as an accepted one does. The test pins it empty, and the
fixture loads **both** stacks first so a lost forward stack is visible at all.

The neighbouring test `a_refused_switch_corrects_the_asker_and_nobody_else`
compared `active_view` alone and stayed green through all of this: the check
that answers a neighbouring question, in the place the doctrine warns about.

---

## The format a projected cell is drawn with (improvise-e0ex)

### a_formulas_format_reaches_its_own_column_and_no_other

`context/code/core.md` recorded, as the more useful half of improvise-711n's
entry, that `ui/project.rs::grid_of` "has no test that renders a cell and
asserts its digits". This is that test, and it is why the gap was worth naming:
the per-formula format is resolved in `improvise-core` and could have been
resolved correctly there while never reaching the grid.

It asserts four things that fail independently: the formula column's cell, the
raw column's cell beside it, and **both** totals. The totals matter separately
because they are formatted on their own path — `GridBody::col_formats`, read off
the top row — so a correct cell says nothing about the strip under it.

Two faults injected:

- `FormatPlan::at` returning the view's format for every key fails the cell
  assertion (`left: Some("0.50") right: Some("0.5000")`).
- Dropping only the `col_formats` write in `viewport::grid_body`, leaving the
  cells correct, fails the totals assertion (`left: "1.50" right: "1.5000"`).

The second is the one that earns the test its length: without it, the totals
strip could round every formula column to the view's format and no assertion in
the workspace would move.

### a_column_total_does_not_inherit_a_row_local_format

Found by review (ChatGPT Codex, PR #125). The first version of `grid_body` read
each column's total format off **row 0**, and `context/code/core.md` argued that
was safe because "where `_Measure` is on the row axis a column is not a measure
at all". That argument is wrong: a formula's target category is *any* category,
so a formula can own a **row-axis item**, and then row 0's format is a row-local
override applied to a total that has summed that row away — and re-ordering the
rows changes the total's spelling.

The fixture builds the same grid twice with `12.5 FORMAT .4` owning the **first**
row in one and the **last** row in the other, and requires the two totals strips
to agree. Every raw cell in `raw_workbook` holds the same value per measure, so
the two grids' totals are numerically identical and the *only* thing that can
differ is the format — the assertion has nowhere to hide.

It asserts the owned cell first (`12.5000`), so a plan that applies no format at
all fails on the fixture rather than passing the comparison vacuously.

Fault injected: `plan.at(layout.cell_key(0, ci).as_ref())` in place of
`layout.column_key(ci)` — the rejected version, restored. Fails with
`left: ["14.5000", "16.5000"] right: ["14.50", "16.50"]`, which is the bug in
one line: row order decided the spelling.