improvise/context/test-evidence/tooling.md

Test evidence — the repository's own gates

Evidence for the checks in scripts/ and for the tests that hold CI's shape. The arguments behind the checkers themselves are in context/code/tooling.md; this file records what fault each test was watched to catch.

Note what is deliberately absent. scripts/check-comment-runs.py --self-test, scripts/kb-test.sh and scripts/assemble-stamp-test.sh are check-the-checker suites whose cases are already injected faults, and none of their case names is a Rust fn or a JS test( the evidence checker could resolve — so their evidence is the case list itself plus context/code/tooling.md, and inventing headings for them here would be headings naming nothing.

"no test in web/tests/ reaches for a built bundle"

The js CI job runs node --test web/tests/*.test.js and installs nothing; the browser job runs web/tests/bundle/*.test.js after three wasm-pack builds. The moment a flat web/tests/ file reads web/pkg/, the fast job inherits the slow one's dependency — and the likely repair is to add a build step to it, silently spending the ~4m30s the split exists to avoid. Until improvise-4ktt that constraint was a directory convention and a glob, checked by nothing, and it failed in the direction that looks like ordinary breakage: a confusing ENOENT in CI on a machine where the bundle happens not to be built.

Fault injected, 2026-09-21: appended const injected = readFileSync("../pkg/improvise_wasm_client.js"); to web/tests/grid_plan.test.js. The test failed naming that file and pointing at web/tests/bundle/. Reverted; green again.

Two things the test does that a simpler one would not. It asserts its own filename is among the files it scans, so it cannot become the one place a violation may hide. And it strips comments before matching, because this file's own header and the test's header both name the directory in prose — a lexical grep over raw text would have had to exempt them, and an exemption is where the next violation goes.

What it cannot see, stated because the check itself uses the evasion: a path assembled at runtime ("../" + "p" + "kg") is invisible to a lexical match, and the test's own FORBIDDEN constant is built exactly that way to stay out of its own scan. That is a deliberate, disclosed hole — closing it needs a parser, and the failure it would catch is one somebody has to write on purpose, where this gate exists for the one written by accident.

"the subdirectory that MAY use a bundle is not the one being checked"

The companion, and the one that stops the first from passing vacuously. If web/tests/bundle/ were emptied or its files stopped touching the bundle, the first test would still be green while having nothing to be scoped against — "no flat test uses the bundle" is uninteresting if no test anywhere does. This asserts the other half of the split exists and really does what the flat half is forbidden to do.

Fault injected, 2026-09-21: moved web/tests/bundle/wasm_surface.test.js out of the tree. The test failed with "web/tests/bundle/ is empty — the split lost its other half" rather than passing. Restored; green again.

"no test in web/tests/ imports a module that reaches the bundle"

The companion to the lexical check above, and the reason that one is not enough: import "../app.js" needs the bundle transitively and names nothing forbidden, so the js job breaks loudly but unattributed. app.js and worker.js are the two entry points that load the bundle; shell.js and serial.js do not, and serial.test.js imports one legitimately.

Fault injected, 2026-09-21: appended import "../app.js"; to web/tests/grid_plan.test.js. The lexical test stayed green; this one failed naming the file. Reverted.

Note — the checker's own blockers, found by review of PR #83

Two gates in scripts/check-comment-runs.py produced a pass while not doing what their prose said. Both were in the mode the same PR rewrote, which is the part worth keeping: writing the doctrine down does not exempt the writer.

A leaf name is not an identity. Six leaf names collide in this workspace and two were cited in evidence. Renaming ui::project::tests::a_load_that_fails_is_not_a_rebase — the one its entry describes — left the checker printing all name a test that runs, exit 0, with the heading answered by the ui::host::tests namesake. Now an ambiguous leaf is refused with both candidates named, and the two entries are qualified. Re-verified both ways, 2026-09-21: the refusal fires on the unqualified headings and names exactly those two; with the heading qualified, the same rename is reported as an orphan rather than absorbed.

--rust-tests was strictly more permissive and said the same words. It called cargo_test_names with one argument, so the --ignored subtraction never ran, and the success line was byte-identical to the honest path's. Injecting ### eyeball_the_rendered_hints_and_pages orphaned on the default path and passed on the flag's. Now the flag requires --rust-tests-ignored (one alone exits 2) and the success line names its source.

A missing evidence directory exited 0. The early return preceded the citation scan, so deleting context/test-evidence/ reported nothing while 170 citations pointed at a directory that was not there. Verified 2026-09-21 by moving the directory aside: exit 1, 170 tests cite context/test-evidence and have no entry there.

Note — the flag that selected the cap away (improvise-ijql)

scripts/check-comment-runs.py --check-evidence early-returned in main, so the flagged run never answered the comment-run cap and printed nothing about having skipped it. The reviewer's position when it was first raised was that CI runs all three spellings (ci.yml, the three check-comment-runs.py steps in the test job) so this was documentation only. The counter that carried is that the flagged form is the expensive one — it needs a compiled workspace — which makes it the one a human runs alone; three agents hit it independently in one campaign, and in one case a local gate went green while CI went red on a three-line doc comment.

Reproduced before the fix, 2026-09-21, in a temporary tree holding one .rs file with a four-line comment run:

$ check-comment-runs.py src --check-evidence --rust-tests … --rust-tests-ignored …
test-evidence: 0 entries, all name a test that runs …
EXIT=0
$ check-comment-runs.py src
1 comment runs longer than 2 lines:
  src/bad.rs:1  4 lines
EXIT=1

Two invocations over one tree, opposite verdicts, and the one that exits 0 is the one a human is more likely to run by itself.

The guard is _dispatch_self_test, which runs the script as a subprocess — the only level at which this is visible, since every unit in the file was correct and the defect was in main's dispatch. Both arms were demonstrated:

injected fault	what failed
restore the --check-evidence early return	DISPATCH: --check-evidence: exited 0 over a four-line comment run
make the cap always report a violation	DISPATCH: bare: refused a clean tree and the same for --check-evidence

The second arm is what stops an always-failing cap from satisfying the first, and it is why SELF_TEST_DISPATCH asserts a clean tree passes as well as a dirty one failing. Why the fix is additive rather than a printed disclaimer, and why --self-test/--list-doctests are excluded from the table rather than exempted from the rule: context/code/tooling.md §2, "A flag adds a question".

Note — the stamp gate's discrimination, measured

scripts/assemble-stamp-test.sh separates three outcomes rather than two, which is what makes it evidence rather than a smoke test. Measured 2026-09-21 over its nine cases: the real script passes 9; deleting the final test -z "$missing" (the PR #69 injection, which the inline version survived at exit 0) passes 4 and fails 5; a script that does nothing passes 2; one that always fails passes 0.

The PR body first cited "4 of 9" for the middle row, from a run taken when the suite had eight cases. A figure is what the next reader checks, so it is recorded here where it can be re-measured rather than only in a message nothing can audit (design-principles §6).

Note — the stamp gate matches literally

scripts/assemble-stamp.sh greps with -F. As a regex, the . in improvise_wasm_client.js?v=abc matches any character, so a stamp corrupted in exactly that position reads as present — a pass in the one direction a guard must not fail in. Inherited from the inline version this was extracted from, and found by reading the extraction rather than by anything failing.

Fault injected, 2026-09-21: dropping the -F makes improvise_wasm_clientXjs?v=abc satisfy the pattern, and a pattern is matched literally, not as a regex goes from pass to fail.

Note — the doc site's derived checks

scripts/build-doc-site.mjs carries two gates that no test names, because the generator is the test: it runs on every assembly and exits non-zero.

Faults injected, 2026-09-21, each observed:

injected	reported
a new context/code/newarea.md	context/code/newarea.md is neither published nor a named omission
context/plan.md moved out of the tree	context/plan.md is listed as omitted but does not exist
a line appended below the generator in web/assemble.sh	web/assemble.sh must invoke this generator LAST, naming the offending command
a trailing echo "see scripts/build-doc-site.mjs"	the same, once the check asks for an INVOCATION rather than a mention — .includes had accepted it (review, PR #83)

scripts/check-ci-graph.py is the third, and it is a script rather than a stanza because its own refusals need checking: adding a zz-new-check: job to ci.yml reports pages-deploy does not need zz-new-check and exits 1, where before the rule was prose and nothing noticed. Its five self-test graphs include the one a naive parse gets wrong — another job's needs: read as the deploy's.

The second arm is the one that would not get written first, and it is the point: without it, deleting or renaming an omitted file leaves an entry asserting a reason for a file nobody can find — the same failure as a dead ticket id, which AGENTS.md already calls worse than no id.

The first arm was filed as latent and was live when it was written: context/code/kb-tool.md had been unpublished, unargued and unnoticed since it was added.

Duplicate evidence headings — SELF_TEST_DUPLICATES

improvise-hsoe. --check-evidence verified that a heading names a test that runs, and a duplicate heading names one just as well, so both copies pass. PR #92 shipped two entries under a_layout_change_moves_the_layout_version_and_no_record, and two more pairs sat in tui.md — one of them asserting that THE_MEMBERSHIP_CORRECT_SPANS was "a live tripwire, do not remove it" some four hundred lines above the entry recording that improvise-gd83 had removed it. A reader meeting the older copy first takes it as current, which is worse than an orphan: an orphan announces itself as broken.

The fault is injected by the six cases in SELF_TEST_DUPLICATES, and three of them are the ones that would not get written first. Two different tests sharing a leaf must come back ambiguous and not duplicate, because the wrong diagnosis files the wrong repair. A prose heading must be free to repeat, because it claims no test. And the bare and module-qualified spellings of one test must collide — grouping by the string rather than by the resolved test would have missed exactly the case improvise-rm4s's qualification repair creates, which is the first way this would have come back.

The live tree was itself the first injection: before the three merges the check exited 1 naming all three pairs with both line numbers; after them it exits 0. The clean arms (two headings naming two tests, and prose repeated) are what stop an always-refuse from satisfying the rest.

The PR review reader's own suite — scripts/pr-reviews-test.sh

improvise-xgra. Like kb-test.sh and pr-watch-test.sh, its cases are already injected faults and none of their names is a Rust fn the evidence checker could resolve, so the case list is the evidence. What is recorded here is the five faults it was watched to fail against, and which case caught each — a suite that only asked "did it print something" passes all five.

Fault 1, the inline surface dropped entirely, which is the live bug as it shipped: inline: [] in both fetch paths. Eight cases fail, including both --wait arms. This is the one that matters, because the old script's behaviour was this fault and it exited 0 on PR #122's review body while printing none of the five rulings underneath it.

Fault 2, an inline comment counted rather than compared: ((.inline|length) > 0) in place of the commit_id == head test. One case fails, an inline comment at an EARLIER head does not satisfy --wait, and that is the pair the two inline fixtures exist to discriminate — the at-head case passes under this fault, so on its own it proves nothing.

Fault 3, the time sort dropped (| . for | sort_by(.at)). Only the three surfaces interleave in time order fails: it reads the three bodies back as SECOND THIRD FIRST, the per-endpoint concatenation order. Every content assertion still passes, because each body is still printed.

Fault 4, no original_line fallback. Only the outdated-comment case fails. line is null exactly when the line has moved under the comment, so a renderer reading .line alone prints draw.rs:null on precisely the comments whose location is hardest to recover by hand.

Fault 5, --wait widened to any issue-level comment. Only an issue-level comment does not satisfy --wait fails. That is the fault a well-meaning repair would introduce — "wait for anyone to say anything" — and the fixture that separates it is a PR thread comment with no commit_id at all.

Two cases are there for the contract rather than the bug: a review body at the current head still satisfies --wait and --wait 0 over an empty PR is success, because other briefs depend on exit 0 and exit 1 meaning what they did before. no PR number is exit 4 pins the third code, which is distinguishable from 1 only if nothing collapses them.

the usage warns about a trailing $? reader asserts a sentence rather than a behaviour, which is unusual and deliberate: the ; echo "exit=$?" trap costs a wrong diagnosis every time it is hit, the warning is the only guard against it, and a header is exactly the thing a later edit trims without noticing.

Three more cases were added after ChatGPT Codex reviewed PR #133, and the pair of bugs it found is the reason the suite has to run in CI rather than only on the author's machine. Both produce exit 2 with no output whatsoever, which is jq refusing an invalid --argjson — so a shell fault at the top of the script arrives wearing jq's exit code, and 17 of 19 cases failed identically with nothing to read.

The first, a multi-assignment local whose third clause reads its first, is fine on macOS's bash 3.2 and fatal on bash 5 under set -u. The suite passed 19/19 locally and 2/19 in CI on the same commit, which is the case for wiring it into the kb job in the same change rather than afterwards.

The second is gh api --paginate emitting one array per page. Fault 6 removes | pages_to_array from the fixture path and exactly the three new cases fail — a comment on the first page survives pagination, …the second page…, and a paginated thread still satisfies --wait (exit 2, the live signature). The fixture carries two concatenated arrays and both paths flatten through the same helper; a fixture holding one array could not have caught this, because it would exercise a shape the live fetch never produces. No PR in this repo has yet had enough inline comments to reach a second page, so the test is the only thing that will ever see it before a user does.

The checker runner's own suite — scripts/checkers-test.sh

The bug behind improvise-f9qf was invisible to a local run, so the evidence that matters here is not "the suite is green" but "the suite goes red when the fault is put back, in the place the fault was not seen."

The fault, injected against the real script. Restoring pr-reviews' multi-assignment local — the exact text 051f541 replaced — and running the suite three ways on one machine in one sitting:

gate	result
sh scripts/pr-reviews-test.sh (the old local gate)	22 passed, 0 failed
direnv exec . ./scripts/pr-reviews-test.sh	22 passed, 0 failed
./scripts/checkers.sh pr-reviews	2 passed, 20 failed, exit 1

The middle row is the one that earns the design. The dev shell supplies the correct bash and would still have shipped this, because pkgs.mkShell exports name and the broken clause read $name. Only scrubbing the environment as well as pinning the shell makes the fault visible, and the 2/20 shape matches the 2/19 CI actually reported.

checkers-test.sh itself is twelve cases, and each injects rather than asserts. The two that carry the argument:

• a suite reading an inherited dev-shell variable fails under the runner — a fixture suite whose body is echo "$name" under set -euo pipefail, run with name=nix-shell-env exported, exactly as direnv leaves it. It is paired with and passes when run directly with that variable exported, which asserts the same fixture passes the old way. Without that pair the first case is a test of set -u and proves nothing about the runner.
• a bash-3.2-only construct fails under the runner's pinned bash — a fixture suite containing local a="$1" b="$a/x", which is the bug's shape reduced to two lines. Green on bash 3.2, fatal on bash 5.

The rest separate the runner's outcomes from each other: a #!/bin/sh script in the directory is refused and the same sandbox passes once it is removed (so the refusal is not standing in for some other complaint); a failing suite fails and a passing one passes; an old bash with no nix is exit 3 rather than a quiet run; an unknown suite name is exit 4 rather than a silent no-op; and --list names exactly the suites present, including the case where an X-test.sh has no X beside it — scripts/node-test.sh is a test runner, and the first version of the glob ran it as a suite, where it printed its usage inside a green gate.

What is asserted about the runner's own environment, and what is not. The allowlist is checked only in the sense that seven real suites pass through it; there is no case asserting a particular name is absent beyond name. A suite that grows a dependency on some other inherited variable will fail loudly on first run rather than be caught here. Stated rather than guarded.

The two-workspace fmt gate — scripts/fmt-workspaces-test.sh

The bug (improvise-b4ks) was a gate decoupled from what it checks while wearing the exit code of a real failure. From the repo root, cargo fmt -p opfs-contract -- --check exits 1 with package 'opfs-contract' is not a member of the workspace — --check never runs. A caller reading $? cannot tell "this crate is misformatted" from "nothing was checked", so the gate could neither report success when the crate was dirty nor failure when it was clean. Reproduced at 7cef9f8, exit 1, before any fix.

scripts/fmt-workspaces.sh replaces it: it discovers every [workspace] root and checks each with -p <member> scoped by that root's own --manifest-path. Exit 2 is a usage error or a cargo that could not run; 1 is a formatting verdict; they are never the same number.

Why -p and not the two obvious spellings — both were measured, and both fail in a direction that looks like success:

• --manifest-path <root> --all crosses the workspace boundary. --all formats "all packages, and also their local path-based dependencies", and harness/opfs-contract depends on improvise-io by path, so the inner root reports the outer tree's files. Measured: an injected fault in src/draw.rs was reported by the harness root's own invocation.
• --manifest-path <root> bare is blind to the members. The repo root is a package (improvise) as well as a workspace, so the bare form checks only that package. Measured: an injected fault in crates/improvise-core/src/format.rs exited 0.

Four rivals were injected into the script and each failed the suite: --all for -p (the boundary leak, caught by counting — the outer file must be named exactly once); the bare manifest-path (caught by the outer-fault case); return 2 collapsed to return 1 on the usage arms (two cases); and "no workspace root found" returning 0 instead of 2.

The fixture had to be made faithful twice, and each vacuity was found only by injecting the rival — not by reading. The suite passed under a rival at both stages, which is the only reason the gaps were visible:

• The first fixture's nested crate had no path dependency into the outer tree, so --all had nothing to leak through and the boundary rival passed.
• The first fixture's outer root was a virtual manifest with no [package], where the bare form happens to reach every member, so that rival passed too.

Both are properties this repo really has, and a fixture that omits them tests a repository that does not exist. That is the generalisable point: a check_the_checker's fixture must reproduce the structural features the hazard depends on, and the only way to know it does is to inject the rival and watch it fail. A green suite over a simplified fixture is the same non-evidence as a green suite over a vacuous assertion.

What is not covered. The script asks cargo metadata for members, so a root whose metadata succeeds but whose member list is wrong is not detectable here. Narrowed by improvise-90ea, below: a list that omits a package in the tree is now refused; one that names a package it should not is still not detected. And harness/opfs-contract's own cargo fmt -p opfs-contract step in the harness-lint job is left as it is: it resolves correctly because that step sets working-directory, and the ticket's own note confirms CI as written was never the broken caller.

The exit contract, narrowed after review — scripts/fmt-workspaces-test.sh

ChatGPT Codex on PR #159 found that the contract above promised more than the script maintained: both formatter invocations end || rc=1, so any failure after the cargo fmt --version probe was reported as "misformatted" even though no verdict had been produced. The finding was correct and was measured rather than assumed:

• an unparseable source (pub fn a( {{{) exits 1 — identical to a diff;
• an unknown rustfmt.toml option is only a warning, and exits 0.

Of the two remedies offered, the contract was narrowed rather than the failure re-classified, because parsing rustfmt's prose to tell a diff from a parse error is precisely the check-adjacent-to-its-subject this repo keeps getting wrong. The line the split actually needs is "did the gate run":

• exit 1 — rustfmt did not certify a root clean. A diff or a refusal to parse; either way a file is named and the tree is what is wrong.
• exit 2 — the gate did not run. Usage, no cargo, no rustfmt, cargo metadata failed, no [workspace] root. No source file is named. Every such cause is checked above the loop, which is what makes the narrowing true rather than hopeful. (Since improvise-90ea, a package no root reaches is also exit 2, and it names that package's Cargo.toml, which is the reason no verdict exists. The line is "no verdict", not "no output"; improvise-ximf.)

Both directions are now pinned, and two further rivals were injected:

• deleting the cargo fmt --version probe;
• turning || rc=1 into || return 2, which collapses a verdict into the usage code — the mirror of the original bug, and caught by three cases.

The first of those exposed a vacuous test, again only by injection. The original stub was a cargo that failed for every argument — but such a cargo also breaks cargo metadata, so the run exited 2 on the metadata arm and the probe could be deleted with the suite still green. A missing rustfmt component is cargo metadata succeeding while only cargo fmt fails, so the stub now delegates every other subcommand to the real cargo and fails fmt alone. Both stubs are kept, because they exercise different arms.

That is the third fixture correction on this one checker, all three found the same way: the suite passed under a rival. The generalisable rule is that a stub must fail in the shape of the real fault, not merely fail — a stub that breaks more than the real thing breaks satisfies the test through the wrong arm, and reads exactly like coverage.

The root package, and why the fourth fix is structural — scripts/fmt-workspaces-test.sh

A post-merge review of PR #159 found a fourth fixture vacuity (improvise-90ea). The fixture's outer root is a package (root-pkg) as well as a workspace, which is right, but no case ever dirtied the root package's own src/lib.rs. Only the member a/ and nested/ were dirtied. So a script that formats every member except the root package passed 17/17. In this repo the root package is improvise, the whole TUI. Reproduced at 40b8913 with the review's rival in members_of (drop the package whose manifest sits at workspace_root): the suite passed 17/17 and --list left improvise out.

Adding one more hand-picked case would have fixed the fourth and not the fifth. Four vacuities in one checker all have the same cause: the cases were chosen by hand, one per hazard somebody had thought of. So the fix is structural, and it has two halves because it has to cover two sets of packages:

• Every package the fixture has. The two hand-picked fault cases are replaced by a loop over every [package] manifest in the fixture. Each one gets its own misformatted src/lib.rs, must exit 1, and must be named exactly once, so the --all boundary check now covers every package rather than one. The enumeration is a grep for [package], deliberately not the script's [workspace] + cargo metadata discovery, since that discovery is what is under test. A count assertion (three packages) stops the loop going vacuous if the fixture loses a package. A package added to the fixture gets its case automatically.
• Every package the tree has. The loop proves only that the script handles the fixture's shapes. A real tree can hold a package no root lists (an excluded crate, say), and nothing in the fixture would say so. So the gate itself now checks an invariant before it formats anything: every [package] manifest the find reaches must be a member of some root, or it exits 2 and names the package. That is exit 2 and not 1 because no verdict was produced for that package, which is what exit 2 means here. Paths are compared through realpath, because on macOS mktemp lives under /var while cargo reports /private/var.

Rivals injected, and where each failed (the suite went from 17 to 21 cases):

• The review's skip-the-root rival. It now fails the fixture at a fault in src/lib.rs is reported (exit 0). With the invariant in place it fails six cases (exit 2, "no workspace root reaches"). It also fails the real repo with no fixture involved: fmt-workspaces.sh exits 2 naming Cargo.toml. That last one is the property worth having. The gate refuses its own blindness in CI, not only in its test suite.
• Disabling the invariant (if false). It fails a package no workspace root reaches is exit 2, not a pass with exit 0, even though that package's src/lib.rs is misformatted. That exit 0 is the hazard: an unreached package is silently clean.
• The two earlier rivals, re-run against the loop. --all for -p fails both a/src/lib.rs and src/lib.rs "named exactly once". The second of those was not predicted: the nested root's --all also reaches the root package. The bare manifest-path form fails a fault in a/src/lib.rs is reported.

The real-repo control from the ticket still holds: a misformatted fn appended to src/ui/text.rs is reported (Diff in …/src/ui/text.rs:72).

What is still not covered. A root whose member list includes something it should not is not detected, and a package in a directory the find excludes (target/, .claude/, node_modules/) is outside the invariant by construction.

The kb tool's completeness guard and atomic write — scripts/kb-test.sh cases 18–22

improvise-usus: scripts/kb add wrote onto a working-tree kb.pl that had lost 276 of its 801 clauses, and kb_consistent passed. The argument for the fix (compare with HEAD by key, record kb rm removals in a ledger, publish by rename) is in context/code/kb-tool.md. This entry covers what each case was watched to catch.

Reproduced at 40b8913 before any fix. A copy of HEAD's kb.pl with 34 of its 35 location/3 facts removed still passes kb_consistent, and kb add onto it printed 775 before, 1 in delta, 776 after and exited 0. Against the unfixed tool, the new cases gave 33 passed, 5 failed: 18, 19, 20b, 21 and 22 failed. Case 20 passed there because an rm followed by an add has no guard to trip.

The fixture has to fail in the incident's shape. The ticket's "truncate to half" fixture also breaks kb_consistent, because the rules live in the second half. So the unfixed tool refused it too, by the wrong guard. Removing a whole predicate is an existence error, and removing cited clauses fails the citation check. Case 18 keeps one location/3 fact and asserts that the fixture passes kb_consistent before using it. Case 19 keeps the ticket's half-truncation, and said() is what makes it evidence: the completeness guard runs first and must be the one that answers.

Rivals injected into scripts/kb, one at a time. Each failed only its own case:

• baseline taken from the file's own pre-write state rather than HEAD: 18, 19 and 20b failed (35/38 passed). This is the guard that would have passed the real incident.
• ledger never consulted: 20 and 20b failed. A legitimate rm then blocks every later write.
• guard stands down whenever the ledger is non-empty: 20b failed.
• candidate built by plain mktemp in $TMPDIR: 22 failed on the directory check.
• candidate copied over kb.pl before the publish point: 22 failed as modified.
• unreadable baseline exiting 1 rather than 3: 21 failed.

A rival can fail wider than its fault. The first form of the "stands down" rival skipped merge complete altogether. That also skipped writing $cand.ledger, so the later mv crashed under set -e and five unrelated cases failed. It proved nothing until it was rewritten to carry the ledger forward.

Checked by hand, not in the suite: the ledger prunes an entry once the key is re-added and once HEAD stops having it (a committed removal).

The kb completeness guard's follow-ups — scripts/kb-test.sh cases 19b, 20b, 23–25

improvise-ih3k, improvise-yea8 and improvise-za97, filed by the review of PR #166. The argument is in context/code/kb-tool.md. This entry covers what each case was watched to catch.

Reproduced at f909cad before any fix. The new cases 23, 23c and 24 gave 40 passed, 3 failed. 23 failed because an add onto a kb that had lost one of two work_done(<fixture id>,bug_fix,…) clauses exited 0. 23c failed because rm of that paired key was refused as "a clause was dropped": its expected total assumed one clause per key. The tickets did not predict that one, and it meant a pair could not be resolved through the tool at all. 24 failed because the stale ledger entry excused the loss.

The ticket's own reproduction defeats a content stamp. The first fix stamped the ledger with the kb's sha256, and case 24 still failed. kb.pl is canonical, so grep -v of one clause from HEAD is byte-identical to what kb rm of that clause writes. The stamp therefore also carries the inode and the nanosecond mtime.

Case 20b had to change, and why that is not a weakening. It made its second loss with sed -i on the kb. Under the fix, any edit outside the tool voids the ledger, so 20b then refused two clauses instead of one and failed on its "1 of" assertion. It was testing the new voiding rather than what its comment names. The second loss now comes from giving the baseline an extra clause (KB_BASELINE). The kb file stays exactly as the tool wrote it, the ledger stays valid, and the case again isolates "excuses only what it names". It also now asserts the unexcused clause is named.

Rivals injected, one at a time. Each failed only the cases listed:

• keys compared as a set (min(1, N) per key) in complete: 23 only.
• rm's expected total back to before - |delta|: 23c's rm step only.
• ledger pruned by key presence rather than count: 23c's final write only.
• ledger honoured whenever it exists: 24 and 25.
• ledger stamped by content alone: 24 only. 25 cannot see it, because its intervening add makes the restored-then-lost file differ in content anyway. 24 is the case that carries identity.
• ledger re-stamped only by rm, keeping the old stamp on other writes: 24b (as named), 20b and 23c. All three write again after an add.
• ledger stamped with the kb as read rather than the candidate: 20, 20b, 23c, 24b and 25. That is wider than any one case, as expected, since no later write can ever match.
• the gh/main scripts/kb (old refusal text, no stamp): 19b, 24 and 25.

Case 25 is the first to run the production paths. Every other case sets KB_FILE, so ledger_path's git rev-parse --git-path branch and baseline_to's git show HEAD were unexercised. It copies the tool into a throwaway git init repo and checks four things. rm records to the git-path ledger. git status shows only kb.pl modified, so the ledger is invisible to git. A later add is excused. After git checkout -- and a hand loss, the add is refused.

Checked by hand, not in the suite: a SIGKILL at the publish point (the KB_TEST_KILL seam) on a copy inside prolog-sessions/ leaves the candidate, its .base, .ledger, .removed and .stamped siblings and .lock. git status --untracked-files=all and git add -A --dry-run both show none of them, and git check-ignore -v names .gitignore:40 (prolog-sessions/*).

The pr-watch mergeability cases — scripts/pr-watch-test.sh

improvise-edpl. The cases are injected faults in themselves. Like the other scripts/ suites, none of their names is a Rust fn, so the case list is the evidence. The reasoning is in context/code/tooling.md, "pr-watch and mergeability".

The bug, reproduced before the fix. Against the script as it stood at ae96b8c, 19 of the 34 new cases failed and all 19 exited 0. UNKNOWN that resolves to CONFLICTING fails is PR #172's shape exactly: every check green, UNKNOWN/UNKNOWN on the first poll, CONFLICTING/DIRTY on the second. The old script broke out of the loop on the first poll and passed. The fixtures are hand-built. They are the subset of the gh pr view --json shape that the script reads: headRefOid, the two mergeability fields, and the rollup's status and conclusion. They are not a captured payload. A stub that errors instead would have exited 4, a wider shape than the fault.

Rival 1: treat UNKNOWN as an answer. The wait condition was replaced with if true. 13 cases fail: every UNKNOWN row of the table, MERGEABLE/UNKNOWN, the timeout case, and UNKNOWN that resolves to MERGEABLE/CLEAN passes. That last one exits 1 under this rival, so the fix cannot pass by refusing everything uncertain. It has to wait. UNKNOWN that resolves to CONFLICTING fails still passes under this rival, because the allowlist refuses UNKNOWN as well. Rival 2 does not separate it either: its first poll is UNKNOWN, which rival 2 still waits on. What separates it is the original fault, which combined both rivals. An earlier version of this entry said rival 2 separated it. That was wrong.

Rival 2: the old verdict. The allowlist was replaced with exit 0 after the MERGEABLE test. Five cases fail: MERGEABLE/BEHIND, DIRTY, DRAFT, SOMETHING_NEW and null. These are the states that look mergeable by the flag and are not. At 128a2a0 it was six, with BLOCKED among them. Since improvise-zlj2, BLOCKED is waited on, so this rival never reaches the verdict for it. The entry below covers that.

a failed check fails without waiting on mergeability runs at --timeout 0. A script that waited on UNKNOWN before reading the rollup would report that known red as exit 3.

The pr-watch needs-gap and pinned-head cases — scripts/pr-watch-test.sh

improvise-zlj2 and improvise-h3ni. These were found by review of PR #186 at 128a2a0, which was ruled do-not-merge.

zlj2: a healthy PR exited 1. The required Browser harness check run exists only after the build job completes (needs: build). In that gap GitHub reports MERGEABLE/BLOCKED while every registered check is green, and the #186 rule refused BLOCKED. The rollup in the new cases is PR #189's, captured verbatim with gh pr view --json. It has thirteen entries with every field GitHub returns. Each gap state drops or rewinds entries of that rollup with jq. Against 128a2a0 three cases failed, all with exit 1:

• the needs-gap before the required check registers is not a failure: the healthy sequence should exit 0.
• a required check that never registers is a timeout: should exit 3.
• table: MERGEABLE/BLOCKED: should exit 3, not 1.

Rival for zlj2: remove MERGEABLE/BLOCKED from the wait. The same three cases fail with exit 1. That rival is 128a2a0's code.

h3ni: nothing pinned the head across the UNKNOWN wait. The case is a head that moves during the UNKNOWN wait is not a pass. Its polls are UNKNOWN@a, then UNKNOWN@b, then CLEAN@b, and it must exit 2. It passed on the first run, so it is a guard for behaviour the script already has, not a reproduction.

Rival for h3ni: re-pin the head whenever mergeable is UNKNOWN. That case alone fails, exiting 0. All 44 others pass. That is the review's point: the older a head that moves mid-wait is not a pass starts from MERGEABLE, so it cannot see this rival.

The pr-watch red-under-BLOCKED and ended-PR cases — scripts/pr-watch-test.sh

improvise-9pfs and improvise-1qaa. Both were found by the re-review of PR #186 at 04230a7.

9pfs: the commonest red had no case. Every red fixture wore UNSTABLE or UNKNOWN. Here, though, a failed required check reports MERGEABLE/BLOCKED, and the table enumerates green rollups only. The new case is a failed required check under MERGEABLE/BLOCKED fails without waiting. It uses PR #189's rollup with Test and lint (host) rewritten to FAILURE, runs at --timeout 0, and expects exit 1. It passed at 202e7eac, since the script already broke on any failed check, so it pins existing behaviour. Rival: the failed-check break gated on state != BLOCKED. That case alone fails, exiting 3 where 1 is wanted, and the other 45 cases pass. Nothing else in the suite could see a refactor that turned every required-check red into a timeout.

1qaa: a merged PR slept the whole timeout. Two case pairs. Each uses the shape GitHub returned live, with green checks. MERGED is UNKNOWN/UNKNOWN (#189 once merged). CLOSED is CONFLICTING/DIRTY (#74). The CLOSED fixture was UNKNOWN/UNKNOWN at b13355c7, and review found that shape unreal. Each pair asserts exit 5, and that the fixture was read once, using the poll counter fetch keeps on disk. Against the script at 202e7eac:

• MERGED: exit 3 after 102 polls within a 5s timeout.
• CLOSED: exit 1 on the first poll. CONFLICTING already ended the wait, but as "not mergeable".

Two rivals:

• The state checked only when the timeout expires. The MERGED read-once case fails (44 polls), and so does the CLOSED exit-code case, because CONFLICTING breaks to exit 1 before the timeout branch is reached. The poll count is what pins "does not wait" for MERGED, which is why each case is a pair.
• Exit 1 instead of 5. Both exit-code cases fail with exit 1. That keeps "already merged" apart from "not mergeable".

29w2: the CLOSED cases could not discriminate. Found reviewing #253 at 8cdea3a4. The only CLOSED shape was #74's CONFLICTING/DIRTY, and that ends the loop on the first poll whether or not the state is checked there. So rival R8 passed 52 of 52. R8 exits 5 for MERGED at the loop top, but for CLOSED only after the loop breaks. Run live against #46 with --timeout 15, R8 exited 3 and 8cdea3a4 exited 5. A third shape is added, #46's: CLOSED, green checks, MERGEABLE/BLOCKED, which the loop waits on. #74's shape stays.

• R8: both new cases fail, the exit-code case with exit 3 and the read-once case after 65 polls. The other 52 pass.
• The state checked only at timeout: the new read-once case fails (77 polls), alongside the MERGED read-once case and the CONFLICTING/DIRTY exit-code case.

The pr-watch ended-PR ordering cases — scripts/pr-watch-test.sh

improvise-24lz, found reviewing #253 at b13355c7. That head checked state after --expect but before pinning the head. So a watch with no --expect that saw head a running, and then head b MERGED, exited 5 where it should have exited 2. It accepted a merge at a head the watcher never saw, which is the PR #67 hazard. Two cases:

• a head that moves and then merges is not reported as merged: head a with a running check, then head b MERGED. Expects 2. At b13355c7 it exited 5.
• a PR merged at a head other than --expect is a head mismatch: head b MERGED, --expect aaaa. Expects 2. It passed at b13355c7, where --expect already came first.

The fix moves the state check below the pin block. On the first poll that block only records the head, so an already-merged PR still ends on its first read. Both new cases pass at 202e7eac, since it never reads state; they pin the order of the new code. Rivals:

• The state checked before --expect and before the pin. Both new cases fail with exit 5, and the other 50 pass.
• The b13355c7 placement. Only the moved-then-merged case fails.

Every JSON parser reads numbers exactly (improvise-dz8c)

every_crate_that_can_parse_json_enables_float_roundtrip

It replaces a check over three crates listed by hand. It reads cargo metadata --no-deps and fails naming every workspace package, and every target-specific entry, whose normal serde_json dependency lacks float_roundtrip. The rule and why it was chosen over the alternatives are in context/code/protocol.md. It is red at 1158c23, naming improvise, improvise-command, improvise-io and improvise-protocol.

Faults injected, each run:

• The feature removed from improvise-io alone. The check fails naming improvise-io (all targets). So did a_legacy_json_workbook_loads_its_numbers_bit_for_bit, until improvise-tt22 removed the legacy load path and that test with it. a_json_import_stores_its_numbers_bit_for_bit stays green, because improvise-command enables the feature and unification reaches io in its build.
• The feature removed from improvise-command alone. The check fails naming improvise-command (all targets). The command crate's own import test stays green: cargo tree -p improvise-command shows io's declaration switching the feature on. So the manifest check is the only guard that sees this case. It is also the case the retired build-graph check would have passed.
• A throwaway workspace member, zz-throwaway, with normal dependencies on improvise-protocol and serde_json and no feature (the lockfile updated, as a real new crate would be). The check fails naming zz-throwaway (all targets). With the unconditional dependency carrying the feature and a cfg(target_arch = "wasm32") one lacking it, the check fails naming zz-throwaway (cfg(target_arch = "wasm32")). That is stricter than the build needs, since a wasm build unifies the two entries, and cheap to satisfy. The crate was then removed.

The checked.len() >= 2 assertion is the positive signal. If a cargo change renamed the metadata fields, the check would find no serde_json dependency and pass vacuously without it.

The re-poll a second boundary could skip — scripts/pr-reviews-test.sh (improvise-qnpi)

What broke. "an inline comment arriving mid-wait is seen" failed once in two identical runs of checkers.sh on a loaded machine. pr-reviews --wait computed deadline = now + WAIT and then tested now < deadline before its first re-poll. With --wait 1, a second boundary falling between those two reads of the clock made the loop run zero times: exit 1 with one poll, whatever the second poll would have shown.

Reproduced under load. A loop running that fixture with one busy loop per core failed once, as exit 1, polls=1, around run 52; the fixed script ran 200 times under the same load with no failure. That is evidence of the mechanism, not proof: the window is a few milliseconds per run.

The case that proves it. "a second passing before the first re-poll does not skip it" puts a date on PATH that advances one second per read (a loaded machine, compressed) and a no-op sleep. Against the old loop it fails every time, exit 1 (22 passed, 1 failed). The fix checks the deadline after each poll, so any --wait above 0 polls at least once more. A larger timeout would not have helped: the case uses the shimmed clock, where only the order of check and poll matters.

The harness lock that every build rewrote — harness/build.sh (improvise-mkbp)

What broke. harness/opfs-contract/Cargo.lock pinned pest 2.9.1 while its path dependencies require pest = "2.9.2", so the committed lock was invalid for its own manifest and every cargo invocation rewrote it.

Why both halves. The resolved lock is committed, and build.sh now starts with cargo metadata --locked against it. The lock alone would repeat the bug at the next workspace dependency bump; --locked alone fails today. The check is its own step because wasm-pack build … -- --locked does not work: with the stale lock restored, that version exited 0 and still rewrote the lock, because wasm-pack re-resolves before the flag reaches cargo build. The standalone check, against the same stale lock, exits 101 in seconds and leaves the file untouched.

The assemble half. web/assemble.sh starts with rm -rf of its output, which is what keeps the deployed tree free of stale files, so it is not taught to spare a harness directory. opfs.spec.js instead asserts /opfs/index.html is served, naming the fix. Injected by running web/assemble.sh harness/serve alone: 4 chromium failures, each carrying that message, where it used to fail with no hint of the cause.

Polling past the second poll — scripts/pr-reviews-test.sh (improvise-7epn)

What was unguarded. Every multi-poll fixture resolved on poll 2, and the timeout case expects exit 1 whether the loop polls twice or twenty times. So replacing the post-poll deadline check with an unconditional break left the suite at 23 passed, 0 failed, and --wait N would have polled at most twice.

Two cases, both on the one-tick-per-read clock. "a review arriving on the third poll is seen" empties polls 1 and 2 and puts an inline comment at the head in last, with --wait 10, and asserts exit 0 and exactly 3 polls. "--wait keeps polling until its deadline" runs the empty fixture with --wait 5 and asserts exit 1 after at least 5 polls. The second one is the timeout path. Its exit code alone cannot tell a loop that stopped early from one that reached the deadline, so it counts polls.

Rival, injected and watched. break in place of [ "$(date +%s)" -lt "$deadline" ] || break: 23 passed, 2 failed. The failures were exactly the two new cases, each "exit 1 after 2 polls". The first attempt to inject used sed -i '', which the GNU sed on this PATH read as a filename. Nothing changed and the suite passed 25/25, which looks like a vacuous rival. The second attempt asserted the replacement landed before running.

Still not covered. Deleting the check outright makes the timeout case loop forever rather than fail. The suite has no wall-clock guard, so that fault shows as a hang, not a FAIL.

A stale harness lock names its remedy — harness/build.sh (improvise-lnsh)

What broke. With a stale harness/opfs-contract/Cargo.lock, build.sh exited 101 and printed only cargo's text, whose help: line says to remove --locked. That is the one change that brings improvise-mkbp back. The real remedy was only in a source comment.

Reproduced, then fixed, by hand. Changing pest's locked version from 2.9.2 to 2.9.1 in the lock reproduces the stale state. Before the fix: exit 101 and cargo's hint only. After: exit 1, and the script's own two lines follow cargo's, telling the reader not to remove --locked and naming cargo update -w --manifest-path harness/opfs-contract/Cargo.toml. The lock was untouched both times. Running that command on the stale lock restored it byte for byte (cmp against the committed file). With the committed lock, build.sh still runs to "harness/serve ready".

Not automated. There is no suite for harness/build.sh. A test would have to corrupt a tracked file and call cargo, which needs the crates.io index, from checkers.sh, whose environment is scrubbed. The by-hand run above is the evidence.

The timeout case asserts the exact poll count, and a hang fails — scripts/pr-reviews-test.sh (improvise-7fl8)

What was unguarded. "--wait keeps polling until its deadline" asserted at least 5 polls for --wait 5. On the one-tick-per-read clock the correct loop polls exactly 6 times: one initial poll, then re-polls until the post-poll clock read reaches the deadline. The 7epn entry above says the poll count tells a loop that stopped early from one that reached the deadline. That was true only for a loop stopping two or more ticks early. Both rivals from the #218 review passed the whole suite at 25/25, re-run here before the change:

• a loop capped at 5 polls whatever the deadline;
• deadline = now + WAIT - 1.

The fix asserts polls = WAIT + 1, which the fake clock makes exact. It also catches an over-polling -le. Against the new assertion both rivals give 24 passed, 1 failed, each "exit 1 after 5 polls, wanted 1 after 6".

The hang, which was wider than the ticket said. 7epn left open that deleting the deadline check hangs the suite. It does not hang only at the fake-clock case. It hangs first at "an inline comment at an EARLIER head does not satisfy --wait", a real-clock --wait 1 case, and at every other case that should time out. So the guard wraps every call. "$REVIEWS" is now a function that runs the script alongside a /bin/sleep 20 watchdog and waits for whichever finishes first (wait -n -p, bash 5.1+). If the watchdog wins, it kills the script and returns 124. /bin/sleep is spelled out because the fake-clock cases put a no-op sleep first on PATH. A first version that killed a watchdog subshell left an orphaned sleep, and could signal a recycled pid.

Rival: the deadline check deleted. The suite now finishes in 93 s with 21 passed and 4 failed. The four timeout cases fail with exit 124; the fake-clock one has made about 400 polls by then. Before the guard the same rival never finished, and an outer timeout 30 killed it at exit 124.

The bash floor is 5.1 — scripts/checkers-test.sh (improvise-fd0j)

What broke. pr-reviews-test.sh calls wait -n -p, and -p arrived in bash 5.1. checkers.sh's find_modern_bash accepted any major version 5. So a bash 5.0 on PATH was pinned, and that suite then failed on an invalid option rather than on anything it tests.

Fix. The probe reads major and minor in one call and requires major*100 + minor >= 501. The exit-3 message and the header say 5.1. context/code/tooling.md states why the floor is 5.1.

Cases (section 5b). A fake bash answering 5 0 with no nix → exit 3. The control, a fake answering 5 1 → exit 0, proves that refusal comes from the version and not from the fake.

Rivals, run against the real suite in a copied sandbox:

• Floor 5.0 (-ge 500) fails "a bash 5.0 with no nix is exit 3, not 5.x" and nothing else.
• The pre-fix runner fails only the 5.1 control. Its probe asked for the major version alone, so the fake's 5 1 was not a number and was refused. The 5.0 case passes against it for the same reason, so that case alone would not have distinguished the old code. The 5.0-floor rival is the one that does.

The dead-ID audit compares positively — scripts/bd-audit-ids-test.sh (improvise-52eq)

What broke. The audit recipe in AGENTS.md ended bd show $(cat ids) 2>&1 | grep -E '^(Issue … not found|…)' || echo "all resolve". It looked for the absence of a negative line. So a bd that was missing, failed, or never ran printed no such line, and the audit reported "all resolve". In this session one attempt at the audit printed "all resolve" because its output file did not exist.

The fix. scripts/bd-audit-ids keeps the file scan and the exclusions as they were. It then takes the known ids from bd list --all --limit 0 --json and comm -23s the cited set against them. It exits 3 if bd is missing, fails, prints no JSON list, or lists no issues. On the tree at 12c81dd it reports 460 cited ids, the same count the old pipeline produced there.

The suite runs a throwaway git repo and a fake bd that answers only list. It has eight cases: a pass, a dead id (exit 1, named), four shapes of a bd that does not answer (no output, [], a failing exit, not on PATH), the real tree against a silent bd, and --help. The passing fixture also cites improvise-core and keeps a dead id in a git-ignored harness/serve/ file. So the pass needs both the exclusions and git's file list.

Rivals, each run:

• The old recipe, pasted in as the script's tail: 1 passed, 7 failed. Every "bd did not answer" case exited 0 with "all resolve", which is the bug. The dead-id case failed only because the fake bd refuses show. That is one more bd that did not answer, so it is not evidence the old recipe misses a dead id against a working bd, which it catches.
• The dead-id comparison disabled (: > for comm -23): only "a dead id is exit 1, and named" failed, with exit 0.
• An empty exclusion list: only the passing case failed, reporting improvise-core as dead.

After each rival the original was restored (cmp against a saved copy) and the suite re-run at 8 passed.

Every checkers.sh suite runs in CI — scripts/check-ci-graph.py (improvise-dak5)

What was unguarded. ci.yml names its checkers.sh suites by hand. A new scripts/X-test.sh beside scripts/X is picked up by checkers.sh locally and never runs in CI unless someone adds it to a step. Found on #226, where bd-audit-ids had to be added by hand.

Planted in the real tree. scripts/zz-planted.sh plus scripts/zz-planted-test.sh appear in checkers.sh --list, and check-ci-graph.py then exits 1 with zz-planted is a checkers.sh suite that no CI step runs; add it to ci.yml. As a control, appending zz-planted to the named step clears it (exit 0, "all 9 checkers.sh suites run in CI"). Both files and the ci.yml edit were then removed, and the real run is back to 8 suites, exit 0.

Seven self-test suite lists. They cover: a suite no step runs; suites named across several steps (one inside nix shell … --command); a bare run covering everything; --self-test covering checkers and nothing else; a bare mention in a YAML comment, which must not count as a bare run; names continued onto the next line with a backslash; and a step naming a suite that does not exist.

Rivals, each run against --self-test:

• YAML comments no longer skipped: exactly the comment case fails, because the bare mention there reads as "everything runs" and the missing suite passes.
• suites_run always answering "everything" (the vacuous shape): five cases fail, every one that expects a suite to be named.

The script was restored after each (cmp against a saved copy), and the self-test is back to "5 graphs and 7 suite lists judged".

The suite check asks checkers.sh what a step runs — scripts/check-ci-graph.py (improvise-maax)

What broke. The adversarial review of #228 at cee61b94, and ChatGPT Codex's P1 on it, found the first version failing open in four shapes. Each let the check pass while a suite never ran:

• a bare scripts/checkers.sh outside a command read as "every suite runs": after an inline # (Codex), in a step name:, or as echo's argument;
• checkers.sh --list alone read as every suite, though --list runs nothing;
• --self-test zz-new kb counted zz-new and kb, though --self-test discards them;
• --list kb zz-new counted both.

Two more were loud but wrong: only the first invocation on a line counted, and quoted names kept their quotes.

The fix moves the argument rules back into checkers.sh. It gains --would-run, which is honoured only as the first argument and reports rather than runs. check-ci-graph.py finds invocations: run: text only, comments stripped by bash's rule, command position only. It then asks --would-run about each one.

Every new case run against the OLD parser (/tmp/ccg-old.py, the file at cee61b94): eight of the sixteen suite cases give the wrong answer there. The five false passes are --list alone, the inline comment (Codex), the non-run block scalar, the step name: and echo's argument. Wrong in a noisier way: --self-test zz-new kb (the old parser counts kb, so it reports only a bogus "zz-new is not a suite"), --list kb zz-new (the same), and a && b (the second suite is missed).

A vacuity fixed before the rivals ran. As first written, the bare-mention cases required zz-new, which no real checkers.sh ever runs. A false bare run then still left exactly zz-new unrun, the expected answer, so the any-word-position rival passed all sixteen. The cases now require kb and pr-watch, and the real step runs only kb. A false bare run is then the only way pr-watch counts as run.

Rivals, one function swapped in at a time and each run against the real --self-test:

• no comment stripping: fails only "a suite named inside a trailing shell comment is not run";
• every line scanned (run: prefixes still recognised): fails "a command-shaped line in a non-run block scalar" and the block-run case;
• checkers.sh accepted in any word position: fails only "a bare mention as echo's argument";
• argv modelled naively (names, else everything): fails the --self-test case and both --list cases;
• first invocation per line only: fails "every invocation on a line counts" and the block-run case;
• --would-run honoured in any position, in checkers.sh itself: fails checkers-test.sh's "--would-run anywhere but first is a usage error" (exit 0, wanted 4) and the parser's "a step spelling --would-run runs nothing".

One more rival crashed instead: an early every-line rival hit shlex's "No escaped character" on a lone trailing backslash. So a line that cannot be parsed now contributes no invocation, which fails closed. Everything was restored (cmp for checkers.sh) and re-run: --self-test reports 16 suite lists and checkers-test.sh 17/17. A planted zz-planted suite still fails the real run, naming it.

The suite gate reads a table, not shell — scripts/check-ci-graph.py (improvise-st8q, improvise-tvte)

What broke, twice over. The #228 re-review at c1417359 showed the maax parser still failing open. On the real ci.yml, two changes each turned exit 1 into exit 0:

• a GITHUB_STEP_SUMMARY heredoc holding a bare ./scripts/checkers.sh;
• folding the kb step as run: >.

That is because run: values were read one line at a time. "--command" in words also matched any command at all. And step and job if: were ignored (tvte). Each round of parsing had been answered by another shell or YAML form.

The structural fix. checkers.sh owns a job→suites table, and each CI job makes one call, checkers.sh --ci <job>. The gate parses no shell. It counts a step only when all of these hold:

• the step sits at step-key indentation inside steps:;
• its whole run: value is that call;
• the call names its own job;
• the step carries no if:, continue-on-error:, shell: or covering defaults:.

Every other line mentioning scripts/checkers.sh is reported. A key with an inline value owns deeper lines, and an open quote owns lines until it closes, so content is never read as structure.

The reviewer's attacks replayed on the real ci.yml (kb step mutated, restored after each):

• the heredoc: exit 1, naming the line and the seven unrun suites;
• run: >: exit 1, the same;
• if: on the kb step: exit 1, "conditional (if)".

Self-test: 19 suite lists. Every case asks the real checkers.sh table.

Rivals, each a source edit loaded and run against the real --self-test, each failing a case only it can:

• No quote tracking: fails "an exact run: line inside a multi-line quoted value does not count" and nothing else. The first quoted case could not isolate this rule, because its content line was not run:-shaped anyway. It was replaced.
• No owner rule: fails "a whole job inside a top-level block scalar does not count".
• Disqualifiers ignored: fails the step if:, job if:, job if: after its steps, continue-on-error and both defaults: cases.
• Any job may run any row: fails "a job running another job's row is refused".
• The call may be embedded in other shell: fails "shell control flow around the call does not count".
• Stray mentions not reported: fails the nine cases that expect a line to be named.
• Steps counted outside steps:: fails "a step-shaped item under a job key other than steps: does not count". The first form of this rival passed, because it changed only where a step starts while the step-key match still required steps:. The rival was rewritten to treat every job key with an empty value as opening steps.

checkers-test.sh gains five --ci cases. The rival that drops "--ci takes a job and nothing else" fails two of them, --ci beside a suite name and beside --self-test, each exit 0 where 4 was wanted.

Stated, not guarded: a step earlier in the same job that rewrites scripts/checkers.sh or its table (context/code/tooling.md).

A continued run: and a skipped need do not count — scripts/check-ci-graph.py (improvise-1be2, improvise-bbwa)

What broke (1be2, P2, fail-open). The gate matched run: against its own physical line only. A plain YAML scalar folds a deeper-indented next line into the value, so a kb step whose run: continued onto a line reading || true ran ./scripts/checkers.sh --ci kb || true. It counted as the exact call, and the gate reported "all 9 suites run". A failing suite could then never fail CI.

What broke (bbwa, P3). The gate disqualified if: on the step and on its own job, but not on a job it needs:. GitHub skips a job whose need was skipped, so if: false on a needed job meant the call never ran, while the gate counted it.

Fix. A run: whose value owns a following non-blank line is not counted, and the step is named. needs: is read in all three forms ([a, b], a scalar, a block list) and walked transitively. A job-level if: anywhere in that closure disqualifies the step as "conditional (needs K, which has if)". The "no CI step runs" message now names the fix site: a row of ci_suites in scripts/checkers.sh, not ci.yml.

Red first. The three new cases (a continuation line, needs: inline and as a block list) each failed with saw [], which is the false pass.

Self-test: 23 suite lists. One of the 23 is the two-hop case, added after the rival below survived the first 22.

Rivals, each a source edit run against the real --self-test:

• The continuation is not marked: fails the 1be2 case only.
• No needs flag: fails the three needs cases.
• Block-list needs: not read: fails "the same, with needs: written as a list" only.
• No transitive walk: passed all 22 cases at first, because every needs case was one hop. "The same, two needs: away" (kb needs mid, mid needs gate) was added, and the rival now fails it alone.

Stated, not guarded: a needed job that fails also skips the job, and an earlier actions/checkout with ref: replaces the tree (context/code/tooling.md).

The gate refuses a key or needs: it cannot read plainly — scripts/check-ci-graph.py (improvise-w0j4)

What broke (P2, fail-open). Found in the delta review of #228 at 2aafff3b. The bbwa fix read needs: names as raw text, and it recognised job and step keys only in the bare key: form. The review put a gate job with if: false in the real ci.yml and gave kb a needs: on it. Every one of these spellings still exited 0 with "all 9 suites run":

• needs: ["gate"], needs: 'gate', and a block item - "gate";
• needs: &d [gate], and needs: &d followed by a block list;
• a flow list split over two lines.

yq read kb as needing gate in every case. The disqualifier keys had the same gap: "if":, 'continue-on-error':, if :, and ? if then : false.

Fix: the review's suggestion, refuse rather than parse. Every structural line inside jobs: must be one plain form. That means a job name, a job key, a step key, a needs: value that is empty, one bare name or a one-line flow list of bare names, or a bare block item. Top-level keys must be plain everywhere. Anything else is reported, naming the line, and the gate fails.

Red first. Eleven refusal cases were added. All eleven failed with saw [] against 2aafff3b, which is the false pass. Two controls, needs: [ gate ] # c and needs: gate, still read gate and are conditional. Two more cases, a quoted job name and a quoted top-level key, were added with the fix.

Replayed on the real ci.yml. Each attack was mutated in, run, and restored, with a byte comparison after the last. All eleven spellings now exit 1, and the plain needs: [gate] control exits 1 as conditional.

The 1be2 boundary. The earlier case put the continuation two columns past the key. So a rival that marks multiline only past owner + 1 passed all 23 cases. On the real ci.yml, a 9-space || true under the kb step's run: at column 8 then exited 0. The new case continues exactly one column past the key. The review's rival fails it and nothing else. On the real ci.yml the 9-space continuation now exits 1.

Rivals, each a source edit run against the real --self-test (39 cases):

• needs: value accepted as anything: fails the five inline or anchored or split needs: cases.
• Block items accepted as anything: fails the quoted-item case.
• Step keys not checked: fails the quoted step key.
• Job keys not checked: fails "if":, 'continue-on-error':, if :, and ?.
• Job names not checked: fails the quoted job name.
• Top-level keys not checked: fails the quoted top-level key.
• Refusals collected but not reported: fails all thirteen.
• The review's owner + 1 rival: fails the boundary case alone.

Stated, not guarded. Content under job keys other than steps: and needs: is not checked. Neither is anything deeper than a step's own keys. The gate reads no structure there.

A job's keys start at indent 4, or its lines are refused — scripts/check-ci-graph.py (improvise-rjaa)

What broke (P2, fail-open). Found in the delta review of #228 at 3eafd4de. w0j4's _readable accepted any line deeper than 4 outside steps: and needs: as content. A needed job whose own keys sat at indent 5 or 6, which is valid YAML, therefore had its if: and needs: read as content. The review's three plants each exited 0 with kb skipped:

• a 6-space gate job, with kb given needs: gate;
• a 5-space gate job, with kb given needs: [gate];
• a 6-space mid job holding needs: gate, with kb given needs: mid.

Fix, as the review suggested. A job's first key must be at indent 4. After that, a line at indent 5 to 8 outside steps: and needs: is content only once its job has a 4-space key. Such a line can only be the content of an empty-valued key, because a key with an inline value already owns every deeper line.

Red first. All three plants, as self-test cases, failed with saw []. A control, env: then FOO: bar at indent 6, stays accepted as content.

Replayed on the real ci.yml. Each plant was mutated in, run, and restored. All three now exit 1, naming the first misplaced line.

Rivals, each run against the real --self-test (43 cases):

• The pre-fix rule (any line past 4 is content): fails the three plants.
• job_keyed never reset at a new job: fails the same three. The previous job's key would otherwise license the next job's misplaced lines.

Refused loudly, not read: an indentless block list, a block list at indent 8, and a quoted top-level key such as "on":. Each is valid YAML, and each is named by line.

A comment on a read line is refused, and the matchers share one string — scripts/check-ci-graph.py (improvise-j49r)

What broke (P2, fail-open). Found in the delta review of #228 at 5738e8e0. _readable checked the comment-stripped line, while _JOB.match in both ci_steps and read_graph ran on the raw line. A header gate: # skipped on purpose was therefore accepted and never opened. Its if: false went to the previous job, and read_graph did not list gate. Planted before wasm: in the real ci.yml, with kb given needs: gate, the gate exited 0: "pages-deploy needs all 7 other jobs; all 9 suites run".

The audit found a second instance. The stripper cut A: "x # y" to A: "x, so _opens_quote held the following lines open, and a job-level if: false after them was never read. Planted under kb in the real ci.yml, this also exited 0 at 5738e8e0.

Fix. A line containing # is refused, with its own message, and the stripper is deleted, so every matcher reads the raw line. read_graph opens a job on _JOB_OPEN ( name: followed by whitespace or the end of the line). That is wider than _JOB, because it refuses nothing. The argument is in context/code/tooling.md.

Red first against 5738e8e0:

• the suite case for the header: saw [];
• the read_graph case for the header: saw [];
• the "x # y" case at indent 10: saw [].

The "x" # note case read correctly at 5738e8e0, because the stripper cut it cleanly. It guards the alternative of reading raw lines without refusing: under that alternative the raw value "x" # note opens the quote tracker. Two old cases were changed. The positive "own row" case lost its trailing comment, and that form is now its own refusal case. The w0j4 padded-list control lost its # c and is still a control.

Replayed on the real ci.yml. Each plant was mutated in, run and restored, and the file was compared byte for byte (sha1 a00288a8…). Four plants were run: j49r as reviewed, j49r with gate also in pages-deploy's needs, "x # y" under kb, and "x" # note under kb. At 5738e8e0 the first and third exited 0. Now all four exit 1 and name the line.

Rivals, each a source edit run against the real --self-test (47 cases). The review's eleven still fail where the review thread said they would. The continuation rival fails two cases, the needs flag five, block-list one, transitive one, flow-only two, owner + 1 one, if_jobs kb-only five, _readable always True fifteen, quoted _NEEDS_VALUE one, and each rjaa rival three. Four rivals are new:

• Raw lines, no refusal: fails the run-line, "x" # note and "x # y" cases.
• The stripped line everywhere, no refusal (the other option): fails all four suite cases.
• Refusal only inside jobs: at indent 8 or less: fails the indent-10 "x # y" case alone.
• read_graph opening on the exact _JOB: fails the read_graph case alone.

The header suite case does not separate the raw-without-refusal rival. There the exact _JOB on the raw line refuses the header anyway, and the complaint has the same text.

A multi-line quoted value is refused, and the quote tracker is gone — scripts/check-ci-graph.py (improvise-6zf1)

What broke (P2, fail-open, predates j49r). Found in the delta review of #228 at 588094b7. The tracker closed a multi-line quoted value at the first line containing its quote character, including an escaped \" or a doubled '', which YAML does not treat as a close. A later content line such as b: " then opened a phantom string that swallowed the job's real if: false. The review planted a gate job whose name: spans lines, with kb given needs: gate and gate added to pages-deploy's needs. yq-go read gate.if == false, and the gate exited 0 in both quote styles.

Fix. A line whose quoted value does not close on its own line is refused, with its own message. The tracker is deleted. The real ci.yml has no multi-line quoted value (scanned with _ANY_KEY and _opens_quote: 0).

Red first against 588094b7: both plant cases (double and single) gave saw [].

New cases, and why each is there.

• The two plants. Each now draws the refusal, and also "conditional (needs gate, which has if)", because with the tracker gone the job's if: is read.
• An opening line ending in an escaped \", and one ending in a doubled '', each followed by needs: ok and a readable closing line (x: y"). yq-go reads each as one string, with kb.needs == "gate". If _opens_quote called either value closed, nothing would be refused, and the gate would read needs: ok in place of kb's real need. These cases are what make _opens_quote's escape handling load-bearing and tested.
• Two refused styles the review listed: a trailing comment on a step, and run: echo "a # b". Both were already refused by the j49r rule, so these cases document behaviour rather than fix a bug. The third style, a quoted with: value containing #, was already the j49r indent-10 case.

Changed cases.

• "an exact run: line inside a multi-line quoted value does not count" is now "a run: line inside a multi-line quoted value: the quote is refused". It draws the refusal of the opening line and a plain refusal of the closing b". Only the exit is guaranteed after a refused line.
• "a closed quote followed by a comment is refused (j49r)" now reads the job-level if: correctly. Its expectation is the refusal plus "conditional (job if)", which is what the j49r evidence wanted but could not have while the tracker existed.

Replayed on the real ci.yml. I planted the review's three shapes: double quote, single quote, and double quote with the reopening line at indent 4. At 588094b7 each exited 0. Now each exits 1, naming the opening line. All seven plants from both rounds were restored byte for byte (sha1 a00288a8…).

Rivals, each a source edit run against the real --self-test (53 cases). All 18 fail:

• the review's eleven;
• the four from j49r;
• three new ones:
  • a multi-line quote is not refused: fails the five 6zf1 cases;
  • _opens_quote ignores \": fails the escaped-\" case alone;
  • _opens_quote ignores '': fails the doubled-'' case alone.

The j49r rival "no comment refusal" now fails four cases, not three: the "x" # note case is also caught by the quote refusal, because the raw value "x" # note does not end in its quote.

Checked, not a finding. A flow value ({…} or […]) that continues onto another line is not refused. yq-go rejects a continuation at a lesser indentation (did not find expected ',' or '}'), which is loud. At the key's own indentation the continuation is read as a key, which fails closed: a spurious if:, or a needs: value the exact _NEEDS_VALUE refuses.

Every value is one of a few one-line shapes — scripts/check-ci-graph.py (improvise-0wn3)

What broke (P2, fail-open). Found in the delta review of #228 at 9dc8a643. A multi-line flow value was not refused:

      - name: [a, "b
        run: nix shell nixpkgs#trealla --command ./scripts/checkers.sh --ci kb
        c: x"]

go-yaml, PyYAML 6.0.3 and Psych 5.2.2 read this step as a single name string with no run:. The gate read the middle line as the step's run: and counted it. The single-quoted form and a flow map did the same. All three exited 0 on the real ci.yml.

This retracts the 6zf1 entry's "Checked, not a finding" note above. That note argued a flow continuation "fails closed", but it only considered closing lines that are not spelled like a step key.

Fix: an allowlist over values, not one more refused construct. Every value the gate reads must be one of: empty; a block-scalar header; a quoted scalar closed on its line; a one-line flow with no quotes, nesting or #; or a plain scalar not starting with a YAML indicator. That covers a key's value, a list item's, or a line of its own. Anything else is refused. The argument is in context/code/tooling.md.

The real ci.yml has 153 plain values, 41 empty, 2 one-line flows and 5 block headers, and none is refused. When a line is both an unreadable key and an unreadable value, the key refusal's message wins, so w0j4's cases keep their diagnosis.

Red first against 9dc8a643. All seven new cases gave saw []:

• the review's three: [ double-quoted, [ single-quoted, {;
• [a, "b], which ends in its closer with its quote still open. This is what the review's proposed fix ("does not close on the same line") would pass if it were read as "ends in ]";
• a flow opened on a line of its own, under an empty name:. yq-go reads it as one string too;
• an anchored flow (&a [x, "b) and a tagged one (!!seq [x, "b), each confirmed with yq-go to swallow the run:.

Replayed on the real ci.yml. I replaced kb's checkers step with each of the review's three plants. At 9dc8a643 all three exit 0. Now all three exit 1, naming the opening line. All ten plants from the four rounds exit 1, and the file was restored byte for byte (sha1 a00288a8…).

Rivals, each a source edit run against the real --self-test (60 cases). All 22 fail. The review's sixteen are my 01–11, the comment rival, the read_graph rival, the old quote-refusal rival (now "value allowlist removed", which fails 12 cases), and both _opens_quote rivals. The four new ones:

• a flow need only end in its closer: fails [a, "b] alone;
• any flow start allowed: fails the five flow cases;
• only a key's value is checked: fails the own-line flow case alone;
• anchors and tags read as plain: fails the anchored and tagged cases.

A valid style newly refused. A quoted key below the _readable depth, such as "FOO": bar under env:, is refused, because the gate cannot tell it from an unclosed quote. Checked directly: refused, and the plain FOO: bar control is accepted.

A line break YAML knows and the gate does not is refused — scripts/check-ci-graph.py (improvise-xpzh)

What broke (P2, fail-open). Found in the delta review of #228 at 7ae871cb. The gate splits on \n only, but YAML also breaks lines at U+0085, U+2028 and U+2029. runs-on: ubuntu-latest<U+2028> if: false on the real kb job is one line to the gate and two to go-yaml. yq-go reads kb.if == false with each of the three characters, and the gate exited 0.

Fix. _UNSAFE_CHAR is checked on every line, before the comment and owner skips. It covers C0 except tab, DEL, all of C1 (which holds U+0085), U+2028 and U+2029. The workflow is read through read_workflow, whose universal newlines fold CR and CRLF. The argument, including why all non-ASCII is not refused, is in context/code/tooling.md. The real ci.yml has 0 CR bytes, no BOM and 27 printable non-ASCII characters, none of them in the refused set.

Red first against 7ae871cb. All five new suite cases gave saw []:

• one per break character, each on a runs-on: line;
• a U+2028 inside a full-line comment;
• a U+2028 inside a run: | body, on the job's last step. yq-go ends the block there and reads the if: as a job key.

Controls. Two cases write a real file and read it back through read_workflow:

• CRLF line ends are accepted with no complaint;
• a lone CR is a line break, so the following if: makes kb conditional, as YAML reads it.

Replayed on the real ci.yml. The three plants, one per character, on kb's runs-on:, exit 0 at 7ae871cb. Now all three exit 1, naming line 255. All 13 plants from the five rounds exit 1. The file was written with newline="" so the characters landed literally, and it was restored byte for byte (sha1 a00288a8…).

Rivals, each a source edit run against the real --self-test (65 suite lists, 2 files). All 27 fail. The 22 from earlier rounds fail as before. Five are new:

Rival	Fails
The character check removed	the five new cases
C1 allowed	the U+0085 case alone
Comments not checked	the comment case alone
Owned content not checked	the block-scalar case, and the comment case
The read does not fold CR (newline="")	both file controls

The owned-content rival also fails the comment case because owner still holds the preceding runs-on: key's column when that line is checked.

Two checks of the harness lock, not one — ci.yml (improvise-dndu)

The CI graph runs two different harness-lock checks, and neither stands in for the other:

• the kb row of ci_suites runs check-harness-lock-test.sh, the self-test. It pins the script's three answers against path-only fixture workspaces. It says the script can tell stale from current from "did not run". It never looks at the real harness lock.
• the test job's step "The harness lock matches the manifests it resolves against" runs check-harness-lock.sh --allow-network against the real harness/opfs-contract manifest. This is the check that actually fails a stale lock.

The recompose of #245 onto #228 dropped the second while keeping the first. The merge resolved the ci.yml conflict with git checkout --theirs, which takes the other side's whole file, not just the conflicting hunk. So the network step, which was not part of the conflict, went with it. Nothing noticed:

• check-ci-graph only asks that every checkers.sh suite has a CI row, and the self-test still had one;
• every other gate is local, and the lock was current.

The merge commit's message even said the step was unchanged. Review and Codex both caught it: git diff 1bec7c0f f41a946b -- .github/workflows/ci.yml was empty, so #245 no longer changed ci.yml at all.

The repair restores the four lines at their original place, after the formatting step. The check that would have caught the loss is the one the repair used: diff the resolved file against each parent, and against the merge base, then confirm every change one side made to the base survives. Against #245's head, the restored ci.yml differs only where #228 moved the kb step's suite list into its table.

A defaults: run: shell: no longer disqualifies a step — scripts/check-ci-graph.py (the user's ruling on #228)

What changed, and why. The gate refused to count any step under a workflow- or job-level defaults: block. The user ruled it "a silly check". It also broke on contact with main: PR #287 added a workflow-level defaults: run: shell: nix develop .#ci --command bash …, and on the merged tree every suite read as run by no CI step. The refusal is removed.

Cases. Two cases pinned the refusal: "a job's defaults: can change the shell, so it does not count" and "workflow-level defaults: disqualifies every step". Both are deleted. The new case "a step under a defaults: run: shell: is counted" carries both a workflow-level default in #287's shape and a job-level one, and expects no complaint.

Red first. Against the code before the removal, the new case drew both steps as "conditional (workflow defaults)" and kb's as "(job defaults, workflow defaults)", so both rows went unrun.

Kept. A step's own shell: still disqualifies it; the ruling did not cover that. The w0j4 case refusing a quoted "defaults": key is about key spelling, and it stands.

The audit tells grep's no-match from its read error — scripts/bd-audit-ids-test.sh (improvise-lqs0)

What broke. bd-audit-ids treated xargs exit 123 as "a grep batch matched nothing". xargs also exits 123 when grep exits 2, so a cited file the scan could not read passed the audit, and any dead id inside it was missed. The usage text promised exit 4 for that. Two smaller mismatches with the stated contract:

• run outside a git repository, the audit exited 1, which is the "dead id" code;
• given a bd list whose objects carry no id, it reported every cited id dead (exit 1) where "bd did not answer" is 3.

Reproduced before the fix, with six new cases: 10 passed, 4 failed.

• a dangling symlink among the cited files: exit 1, the read error ignored;
• an unreadable (chmod 000) file holding the only dead id: exit 0, "all resolve", which is the false pass;
• outside a repository: exit 1;
• id-less JSON: exit 1.

The control (both files readable, dead id found: exit 1) and the deleted-file case passed before and after.

The fix:

• Each grep batch runs through bash -c and maps exit 1 to success and exit 2 to 255. xargs stops on 255 with 124, so a read error is exit 4.
• A tracked file deleted from the working tree is filtered out on purpose. Its ids are leaving the tree, which is the case the ticket asked to decide rather than leave to accident. A dangling symlink is still an entry, so it still counts as a read error.
• Outside a repository is exit 4.
• bd's output must be a JSON array whose every element has a string id, or the audit is exit 3.

Rivals, each run against the suite (the script restored and checked with cmp after each):

• read error mapped to success (-le 2): fails both read-error cases, 12 passed and 2 failed;
• the deleted-file filter removed: fails "a tracked file deleted from the working tree is not scanned" with exit 4. So the decision is pinned, not incidental;
• the old cd "$(git rev-parse --show-toplevel)": fails "run outside a git repository is exit 4" with exit 1. A first attempt at this rival still exited 4 and proved nothing; it was replaced;
• the id check reduced to type == "array": fails the id-less case with exit 1.

The chmod 000 case is skipped when the suite runs as root, where the file stays readable. The dangling-symlink case covers the same path for every user.

Absent is not deleted — scripts/bd-audit-ids-test.sh (improvise-tdfd)

What broke. The file filter [ -e "$f" ] || [ -L "$f" ] dropped every tracked path missing from the tree, and took each one to be a deletion. A path marked --skip-worktree (or left out by a sparse checkout) is absent too, but it is not deleted. Its ids were never read, and the audit passed. ChatGPT Codex found this on PR #244. The adversarial review reproduced it with context/adr/001-concurrent-edits.md, the only citation of one id: exit 1 → 0.

Fix, Codex's suggestion. An absent path is skipped only when git ls-files -d lists it. Any other absent tracked path is exit 4, and the path is named. The xargs comment no longer claims 124. That is GNU; BSD exits

1. The script tests nonzero, so both work.

Red first. "a skip-worktree file missing from the tree is exit 4, and named" exited 0 against the pre-fix script, reporting "all 1 cited ids resolve".

Rival: every absent path reported as unread (the ls-files -d test removed). It fails "a tracked file deleted from the working tree is not scanned" alone, so the deleted-file case is what stops the fix from over-refusing. The rival ran from an ignored copy inside the worktree. A copy under /tmp fails "the real tree does not pass against a silent bd" for a reason unrelated to the rival: there is no repository there.

Deleted means git will record it — scripts/bd-audit-ids-test.sh (improvise-cota)

What broke. The tdfd fix above asked git ls-files -d whether an absent path was deleted. ls-files -d also lists an absent path marked --assume-unchanged. But git status shows nothing for that path, and commit -a would never record the deletion, so its citations stay in HEAD. The audit skipped the file and exited 0. This is the tdfd shape again: absent, not a deletion git will record, and never read. It was found in the delta review of PR #244 at 33d563a9.

Fix, the reviewer's oracle. A path counts as deleted only when git diff --name-only --diff-filter=D lists it. That compares the index with the working tree and honours both index flags: it hides a skip-worktree path and an assume-unchanged path alike, and lists a plain unstaged delete. A staged git rm never reaches the question, because the path has already left the index that ls-files --cached enumerates.

Red first. "an assume-unchanged file missing from the tree is exit 4, and named" exited 0 against 33d563a9, reporting "all 1 cited ids resolve".

Rivals.

• The ls-files -d oracle, which is the red state above.
• Every absent path reported as unread (elif true). It fails "a tracked file deleted from the working tree is not scanned" and nothing else, so the deliberate skip of a real deletion survives the stricter oracle.

This rival was edited in place in the worktree script, not copied to /tmp, for the reason the tdfd entry gives.

Section 5b runs on a hermetic PATH — scripts/checkers-test.sh

What broke. Every PR's "Knowledge base consistency" job went red at 666de199 with "a bash 5.0 with no nix is exit 3, not 5.x (exit 1, wanted 3)" and "a bash 5.1 is accepted (exit 1, wanted 0)". Neither PR alone was wrong. The NixOS-runner change (af44653f) found that /usr/bin:/bin on NixOS holds only env and sh, so the runner dies on dirname with exit 1. It fixed section 5 by moving it to a utilbin of linked tools. #237 (e45a9c42) added section 5b on its own branch, with the old PATH=$work/fakebin:/usr/bin:/bin. The two merged with no conflict, and each had passed CI only on its own side.

Reproduced on macOS by running both 5b invocations with /usr/bin:/bin replaced by a directory holding only env and sh. Both exited 1 on dirname: command not found. Those are the CI exits.

Fix. 5b uses utilbin as well. utilbin gains head and env, because the 5.1 control runs past the version check into check_shebangs and run_suite.

Rivals, each run in the suite:

• dirname left out of utilbin fails all three section-5 cases with the exact CI pair: exit 1 where 3 was wanted, and exit 1 where 0 was wanted. The NixOS failure now reproduces on any host, because no case reads the host's /usr/bin any more.
• env left out fails "a bash 5.1 is accepted" alone.
• head left out fails nothing. check_shebangs calls head inside check_shebangs || rc=1, where set -e is off. A missing head makes every first line empty, and the shebang check then passes without checking anything. That is a defect in checkers.sh and is not fixed here.

The pr-reviews watchdog resolves its tools from PATH — scripts/pr-reviews-test.sh (improvise-3uow)

What broke. The suite hard-coded /bin/sleep for the watchdog in bounded_reviews and /bin/date for the fake clock's pass-through. On the NixOS runners /bin holds only sh. With no /bin/sleep, the watchdog exits at once. wait -n then sees it finish before pr-reviews, and every bounded call reports 124 ("killed after 20s"). CI never showed this, because the kb job stopped one step earlier on the section-5b failure (#296).

Reproduced by pointing both paths at /nonexistent/. The result was 0 passed and 25 failed, every --wait case exiting 124: the same shape the reviewer simulated.

Fix. REAL_SLEEP comes from type -P, before any case puts its stub sleep first on PATH. type -P returns a path even if sleep has been loaded as a builtin. The fake clock's /bin/date pass-through was deleted rather than re-pointed. pr-reviews only ever calls date +%s, so the branch never ran (found by Copilot on #297). The fake now refuses any other arguments with exit 2, so a new date call in pr-reviews shows up as a failure instead of untested code.

Rivals:

• REAL_SLEEP=/nonexistent/sleep fails the whole suite, 0 of 25, so the resolved path is the one the watchdog runs.
• pr-reviews changed to call date -u +%s fails "a review arriving on the third poll is seen" and "--wait keeps polling until its deadline" (23 of 25 pass). The fake's refusal is therefore visible, not swallowed.

Scope sweep. grep -rnE '/(usr/)?bin/[a-z]' over scripts/, web/*.sh, harness/*.sh, harness/tests and .github/workflows finds only #!/bin/sh shebangs, and /bin/sh exists on NixOS. It also found a ${PATH:-/usr/bin:/bin} fallback in checkers.sh, which is now removed (next entry).

The bash floor and its self-tests are retired — scripts/checkers.sh

The ruling. The user ruled that CI should not care which version of bash or of the standard utilities it runs, and that CI jobs should use whatever the .#ci dev shell provides. So checkers.sh no longer pins a bash: find_modern_bash, the 5.1 floor (improvise-fd0j) and exit 3 are gone. Suites run under $BASH, the bash running the runner, with PATH unchanged. The reasoning is in context/code/tooling.md §5.

Tests removed. Three sections of checkers-test.sh were removed. They existed only to pin the floor, and the entries above that cite them describe cases that no longer exist:

• Section 2, "a bash-3.2-only construct fails under the runner's pinned bash". It asserted a bash-5 behaviour, so its result depended on the bash version.
• Section 5, "an old bash with no nix is exit 3".
• Section 5b, "a bash 5.0 with no nix is exit 3, not 5.x" and "a bash 5.1 is accepted", together with the utilbin that #296 added for them.

The suite keeps 10 cases: environment scrubbing (with its direct-run pair), the shebang rule (with its control), fail and pass, unknown suite, and --list.

Kept deliberately. The environment allowlist. That is what the set -u/$name evidence above earns, and it does not depend on any version.

Also removed with the ruling. scripts/kb's stat -f fallback: kb already requires the dev shell for tpl, and the dev shell's stat is GNU. From ci.yml, the toolchain version probes ("the dev shell supplies the pinned web toolchain", "Playwright matches the dev shell's browsers") and every shell: bash override, so every step runs in nix develop .#ci.

A thread past ARG_MAX is read from stdin — scripts/pr-reviews-test.sh (improvise-5jxt)

What broke. fetch handed the review, inline-comment and issue-comment JSON to one jq -n --argjson … call as arguments. Once #210's comments outgrew the kernel's argument limit, execve refused to start jq: jq: Argument list too long, exit 126, with nothing printed. That happened with or without --wait. The script is how agents read reviews, so an ADR steward was blocked on it.

Fix. A combine helper sends the three documents to jq -s on stdin through printf. printf is a bash builtin, so no execve sees the data. Both paths use combine: the fixture path and the live gh path. The other jq calls in pr-reviews already read $json from a pipe. --arg head carries only a 40-character SHA.

The fixture. Each of the three surfaces carries about 2 MB: a review body, an inline comment followed by a small one after it, and an issue comment. 2 MB exceeds both macOS's total ARG_MAX (1 MB) and Linux's per-argument MAX_ARG_STRLEN (128 KB). Every surface is large, so a fix that moves only one document off argv still fails. The file is built with head -c, tr and printf, so it depends on no tool version and no path.

Red first. Before the fix, "a thread larger than ARG_MAX is still read" and "… still satisfies --wait" exited 126, as #210 did live.

Rivals (each 25 passed, 5 failed; the five are the new cases):

• The pre-fix script, restored from gh/main: exit 126 on both run cases, and all three print assertions miss.
• A partial fix that pipes the review and inline documents but still passes the issue comments with --argjson issue "$3": also exit 126. So the large issue fixture is load-bearing, not decoration.

Live. ./scripts/pr-reviews 210 exits 0 with its output sent to a file. The output includes the user's inline answers on context/adr/007-stable-identities.md from 2026-09-24T19:50:22Z onward ("full unicode", "Just a plane number.", …).

Scope sweep. scripts/pr-watch passes its JSON through printf '%s' | jq, never as an argument. grep -rn 'argjson\|--arg ' over scripts/, .agents/ and .claude/skills/ finds no other data argument.

An agent authors its commits, and the committer is left alone — scripts/agent-git-test.sh (improvise-1xfb.3)

What was missing. Every agent commit was authored as the repository owner, so history could not say which agent wrote what. scripts/agent-git runs one git command with GIT_AUTHOR_NAME="Claude ($BEADS_ACTOR)" and refuses, exit 2, when BEADS_ACTOR is unset or empty.

Why not git -c user.name=…, the mechanism the ticket and the brief both offered first. user.name is the default for the committer as well as the author, so it renames both. The ticket asks for the committer to be left alone. Probed in a scratch repo: a -c user.name commit reads Claude (u) <o@x> | Claude (u) as author and committer. A GIT_AUTHOR_NAME commit reads Claude (t) <o@x> | Owner. The same probe showed that a merge takes GIT_AUTHOR_NAME too, and that a plain --amend keeps the author it amends.

A cherry-pick keeps its source's author (Codex P2 and the review on PR #310). The first version of the rule listed cherry-pick among the wrapper's subcommands, but git ignores GIT_AUTHOR_NAME there. That is the behaviour this repo wants, since rule 8 allows only cherry-pick -x of commits that keep their own author. So the contract names only commit and merge. The case "a cherry-pick keeps its source's author" picks a commit authored Claude (agent-b2) with the wrapper set to agent-a1, and asserts agent-b2. A pick that failed would leave the owner-authored base at HEAD and fail the case too.

Rivals (run against the first eleven cases).

• exec git -c user.name="Claude ($BEADS_ACTOR)" "$@": 10 passed, 1 failed, "the committer is untouched" (got Claude (agent-a1)).
• The refusal removed (if false): 7 passed, 4 failed. They are "an unset BEADS_ACTOR is refused" (exit 1 from set -u, not 2), "the refusal names BEADS_ACTOR", "an empty BEADS_ACTOR is refused" (exit 0: the commit landed under the name Claude ()), and "a refusal commits nothing". So set -u alone would not cover the empty case.

Fixture. The repo's identity comes from its own config, with GIT_CONFIG_GLOBAL=/dev/null and GIT_CONFIG_NOSYSTEM=1, and every GIT_* identity variable is unset first. So a caller's own identity cannot make the author or committer assertions agree by accident.