Generalizes the no-session-narrative rule from docs/handoffs/chicorytv-issue-queue.md (where it was correct but scoped to one file, for a file-specific reason) to every durable artifact, with the destination and the carve-out both stated.
What ships
docs/decisions/records/docs/no-session-narrative.md (docs.no-session-narrative) — a doc records the END STATE; the path to it goes in the commit message, the issue, or the ## Closing record. Concretely: a review finding is answered in the commit message and only the corrected claim enters the doc. Carries the who-benefits test and the carve-out (dated measurements, stated snapshot boundaries, tested-and-rejected negative results, the why behind a non-obvious choice, a trap and its consequence), plus the wholesale exemption for docs/decisions/**.
Kickoff file now cites the general rule and keeps only its extra file-specific cost on top, instead of restating the rule.
scripts/check-doc-narrative.sh + a docs-reminder step — ADVISORY. Exits 0 on every path; emits ::warning:: only. --diff <base> scans lines ADDED in the PR (the CI mode); --all sweeps the tracked corpus.
Rows in docs/guard-inventory.md (TOOLING) and docs/remote-state-inventory.md (N/A), plus a docs/README.md task-signal row.
Why advisory and not a gate
A narrative detector is a string predicate over prose. docs/defect-shapes-773.md §4 and testing.guard-derives-population-from-source both argue that class must not be load-bearing, and the withdrawn test_review_verdict_vocabulary_parity.py (six rounds, then deleted) is the empirical case. The record says in as many words that this may never become a blocking gate.
Evidence it is a weak predicate, measured rather than asserted: on the same corpus a reading sweep found 18 sites and this detector found 10 — a union of 21, with each side missing sites the other caught.
Verification
scripts/tests/ — 771 passed, 2 skipped.
Both inventory guards were witnessed red before their rows existed (test_the_inventory_covers_exactly_the_guards_that_exist and test_every_in_scope_file_has_a_row_and_every_row_names_a_real_file each named the new script).
Positive control for --diff: a throwaway commit adding a narrative line to docs/testing.md, a matching line to a decision record, and one to the script itself → flagged exactly the docs/testing.md line, at the correct line number, exempted the record and the script, exit 0.
Negative control: this PR's own diff → nothing to flag, exit 0.
decisions_validate.py, build_decisions_catalog.py --check, check-kickoff-guard.sh all green.
Deferred, stated rather than smuggled
The sweep reports and does not remediate. 21 existing sites across 6 docs are listed on #784; each needs the who-benefits judgement applied by someone who can see whether the prior-state note is load-bearing (several in docs/remote-state-inventory.md are — "graded down by cold review" tells the next reader not to regrade it back), and this repo has the inverse failure on record (#542, pruning the only copy). Follow-up filed.
Closes #784.
Generalizes the no-session-narrative rule from `docs/handoffs/chicorytv-issue-queue.md` (where it was correct but scoped to one file, for a file-specific reason) to every durable artifact, with the destination and the carve-out both stated.
## What ships
- **`docs/decisions/records/docs/no-session-narrative.md`** (`docs.no-session-narrative`) — a doc records the END STATE; the path to it goes in the commit message, the issue, or the `## Closing record`. Concretely: a review finding is answered in the commit message and only the corrected claim enters the doc. Carries the **who-benefits test** and the **carve-out** (dated measurements, stated snapshot boundaries, tested-and-rejected negative results, the why behind a non-obvious choice, a trap and its consequence), plus the wholesale exemption for `docs/decisions/**`.
- **Kickoff file** now cites the general rule and keeps only its *extra* file-specific cost on top, instead of restating the rule.
- **`scripts/check-doc-narrative.sh` + a `docs-reminder` step** — ADVISORY. Exits 0 on every path; emits `::warning::` only. `--diff <base>` scans lines ADDED in the PR (the CI mode); `--all` sweeps the tracked corpus.
- Rows in `docs/guard-inventory.md` (TOOLING) and `docs/remote-state-inventory.md` (`N/A`), plus a `docs/README.md` task-signal row.
## Why advisory and not a gate
A narrative detector is a string predicate over prose. `docs/defect-shapes-773.md` §4 and `testing.guard-derives-population-from-source` both argue that class must not be load-bearing, and the withdrawn `test_review_verdict_vocabulary_parity.py` (six rounds, then deleted) is the empirical case. The record says in as many words that this may never become a blocking gate.
Evidence it is a weak predicate, measured rather than asserted: on the same corpus a **reading** sweep found 18 sites and this **detector** found 10 — a union of 21, with each side missing sites the other caught.
## Verification
- `scripts/tests/` — 771 passed, 2 skipped.
- Both inventory guards were **witnessed red** before their rows existed (`test_the_inventory_covers_exactly_the_guards_that_exist` and `test_every_in_scope_file_has_a_row_and_every_row_names_a_real_file` each named the new script).
- **Positive control** for `--diff`: a throwaway commit adding a narrative line to `docs/testing.md`, a matching line to a decision record, and one to the script itself → flagged exactly the `docs/testing.md` line, at the correct line number, exempted the record and the script, exit 0.
- **Negative control**: this PR's own diff → `nothing to flag`, exit 0.
- `decisions_validate.py`, `build_decisions_catalog.py --check`, `check-kickoff-guard.sh` all green.
## Deferred, stated rather than smuggled
The sweep **reports** and does not remediate. 21 existing sites across 6 docs are listed on #784; each needs the who-benefits judgement applied by someone who can see whether the prior-state note is load-bearing (several in `docs/remote-state-inventory.md` are — "graded down by cold review" tells the next reader not to regrade it back), and this repo has the inverse failure on record (#542, pruning the only copy). Follow-up filed.
The rule existed, correctly stated, scoped to one file: a paragraph of
narrative in the kickoff handoff is a tax paid by every session that pastes
it. That reason is file-specific. The general one is broader and applies to
every doc: a reader coming cold never saw the earlier draft, so "we
previously got this wrong" carries nothing they can act on.
- docs/decisions/records/docs/no-session-narrative.md — the rule, the
who-benefits test, the named destination (the commit message), and the
carve-out that keeps it from over-applying.
- The kickoff file now cites the general rule and keeps only its extra,
file-specific cost on top.
- scripts/check-doc-narrative.sh + a docs-reminder step — ADVISORY. It exits
0 on every path. A narrative detector is a string predicate over prose, and
defect-shapes-773 §4 argues that class must never be load-bearing.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Both inventory guards were witnessed RED before these rows existed:
test_the_inventory_covers_exactly_the_guards_that_exist and
test_every_in_scope_file_has_a_row_and_every_row_names_a_real_file each
named scripts/check-doc-narrative.sh as unclassified.
The `tail -n +5` on `git diff` assumed a four-line file header, which is
wrong for a rename or a mode change; the extra header lines would then be
counted as content until the first `@@` reset the counter. Match the header
lines explicitly instead, `+++`/`---` before the `+*`/`-*` arms.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Cold review (BLOCKED) falsified three of the four things the PR asked it to
attack. Reproduced each before acting:
B1 the script could exit NON-ZERO on two reachable paths, against an
invariant asserted absolutely in four places
H1 an unresolvable base ref printed "nothing to flag" and exited 0, having
scanned zero files — indistinguishable from a clean run
M1 `'\\ No newline'*` never matched (two literal backslashes inside single
quotes; git emits one), so the marker counted as content
M2 an added line whose own text began `++ ` was eaten by the `+++ ` header
arm. Verified against the real predecessor: it reported line 2 where the
truth is 3.
M3 core.quotePath hid non-ASCII paths from BOTH modes, silently
M4 --all dropped a final unterminated line
Those are one mistake at six sites: deciding what a diff line IS from its
prefix, with no hunk state. Patching them individually is the shape that does
not converge here, so the mechanism is replaced — check-doc-narrative.py
parses with hunk state, uses core.quotePath=false, keeps rename detection, and
returns 0 from a bare `except` as the last line of defence.
H2 said TOOLING/NONE is the classification under which no proof is ever owed,
and that a 99-line parser shipped with four demonstrated defects a fixture
directory would have caught. Correct: scripts/tests/test_check_doc_narrative.py
now asserts exit 0 per argument shape, compares every reported line number
against truth read back off disk (never a hand-written number), covers each
defect above, and carries a mutation proof that disarms PATTERNS so the suite
cannot pass on a detector that finds nothing.
M5/M6: docs/ci-cd.md now describes both steps of the job, and the record's
mechanics: states the detector's population instead of implying it matches the
rule's — skills and nested markdown outside docs/ are out of scope.
Proves: scripts/tests/test_check_doc_narrative.py
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Re-review confirmed all seven earlier findings fixed and blocked on three new
ones. Each reproduced before acting.
H3 the never-fails invariant exited 1 on a clean ASCII repo: the summary line
and the bare-except fallback both carry U+2014, so under ascii/latin-1
stdio the code guaranteeing exit 0 was itself what raised. stdout/stderr
are reconfigured with errors="replace" at import. Not reachable on the CI
image (PEP 540), but the invariant is asserted without qualification in
four places, and the earlier B1 was upheld on exactly that standard.
H2 four shipped references to the deleted .sh remained — including the
record's own ENFORCEMENT clause and the generated catalog a lookup
resolves to. A record naming a path that does not exist at its own head is
worse than no record.
M5 a stray `diff.dstPrefix` produced `scanned 0 file(s)` on a file that
plainly contained a hit. Pinned --src-prefix/--dst-prefix/--no-ext-diff:
closing the channel, not the one variant that was demonstrated.
M4 python3 is not guaranteed on the bare `small` lane — docs/ci-cd.md says so
36 lines below the section this PR rewrote, and every other python-using
job there declares setup-python. A missing interpreter is exit 127, a RED
advisory job, which is the one thing this step must never be.
Then I mutation-checked my own new tests, and two of them were vacuous:
- `--diff-filter=d` was a SECOND mechanism for excluding deletions, and
neither it nor the `+++ /dev/null` arm can be witnessed failing, because a
deletion contributes no added lines either way. Removed the duplicate and
labelled the survivor defensive rather than load-bearing.
- the rename test restated a git default. `diff.renames` is true by default,
so it passed with or without --find-renames. It now sets the config to
false, which is what makes that flag load-bearing and the test a proof.
Every remaining load-bearing clause was disarmed individually and witnessed
red: core.quotePath, the hunk-state parse, the no-newline arm, splitlines,
the SCANNED-NOTHING branch, the stdio reconfigure, the prefix pin and
--find-renames.
The previous commit's `Proves:` trailer sat in its own paragraph, so git
parsed no such trailer and prove-fix reported that nothing was verified. It
belongs in the same block as the other trailers.
Proves: scripts/tests/test_check_doc_narrative.py
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Round 3. Both the reviewer and I independently found the SAME defect —
`color.diff=always` / `color.ui=always` injects ANSI escapes and every parse
arm misses, so a real hit reports `scanned 0 file(s)`. That is the THIRD
gitconfig knob in three rounds to turn a hit into a clean-looking run
(core.quotePath, then diff.dstPrefix, now color.*).
Three instances of one mechanism is the point where naming a fourth knob is
the wrong move, so the channel is closed rather than the variant: the user's
and system's config files are removed from the picture via GIT_CONFIG_GLOBAL/
SYSTEM/NOSYSTEM, which also covers knobs nobody has named; every knob that IS
named is pinned on the command line where it beats a repo-local config too;
and each has a row in FORMAT_KNOBS in the test file so the next one found gets
a row instead of a fourth round. The now-redundant `color.ui=never` -c pin was
dropped — a second mechanism on a condition another mechanism already proves
is the duplicate-guard shape.
H-2 was a docstring of mine asserting a proof the script's own comment says
does not exist, and the reviewer showed it is false: removing the
`+++ /dev/null` arm reddens nothing. The arm is defensive; both places now say
so. A docstring claiming a proof it lacks is what a later reader trusts
instead of re-checking.
Eight clauses were load-bearing and unwitnessed. Each is now disarmed and
witnessed red: --no-color, --no-ext-diff, subprocess errors="replace" (a UTF-8
doc under an ascii locale silently emptied the whole scan while exiting 0),
the `diff --git` in_hunk reset (every file after the first dropped from a
multi-file diff — which is every real CI run), the SCANNED-NOTHING return, the
population count, re.IGNORECASE, and the core.quotePath pin.
Two of my own assertions were vacuous and are fixed rather than kept:
`assert "nothing to flag" not in stdout` named a string that occurs nowhere in
this script, so it could never fail; and the LC_ALL=C stdio case passed
against the pre-fix script too (PEP 540), reading as a second witness where
there was one.
M-6: the setup-python step added last round could itself redden the advisory
job. `continue-on-error` on both steps — a script returning 0 does not keep a
JOB green.
The record now records the unhandled-exception arm as UNPROVEN rather than
implying the whole invariant is tested.
Proves: scripts/tests/test_check_doc_narrative.py
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Round 4 confirmed the git-format channel is genuinely closed — an independent
46-configuration matrix across two modes found zero false cleans, and all
eight claimed mutant kills reproduced. What it blocked on was three sentences,
in a PR whose whole subject is that a doc asserting a proof it lacks is what a
later reader trusts instead of re-checking. Mine did it three times.
M-C the workflow comment named "a checkout hiccup" as the motivating failure
and then left Checkout unguarded, so its closing claim was false of the
code it annotated. Checkout is now stated as deliberately OUT of scope: a
job that cannot run is a different failure from an advisory one that ran
and disagreed. The comment also records what I measured on the real
runner (PR#811, run 2179) rather than what I assumed — both steps green
under continue-on-error, job success, commit status success.
M-F `# -U0 is what the line arithmetic assumes` is false: removing -U0
reddens nothing, because the context arm handles any -U. The row pins
that a configured diff.context cannot beat the CLI, and now says so.
N-E `mechanics:` claimed EVERY knob that can reshape the output has a row.
Six knobs do not — they are harmless, which is why. Worse, the claim is
unprovable in principle, which is the exact argument this fix used to
close the channel instead of enumerating it.
And the reviewer's larger point, taken: rounds 3 and 4 each converted "N
load-bearing clauses have no mutant" into "N kills", and round 4 produced six
more from the same well. That is two rounds of one mechanism, so the answer is
to stop enumerating clauses, not to add five more tests. The record now
carries the COVERAGE BOUNDARY in one sentence — the parse path and the format
pins are mutation-proven; the error and degradation arms are defensive and
unproven — and the two inventories point at it instead of repeating a flat
"exits 0 on every path".
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The boundary sentence I wrote last round said "the PARSE path and the FORMAT
PINS carry mutation proofs". Two format pins do not: `-U0` (the context arm
handles any -U, so removing it reddens nothing) and the GIT_CONFIG_GLOBAL/
SYSTEM/NOSYSTEM overrides — which the script's own comment already disclaims
in as many words, so the record asserted a proof the file it describes denies,
two files apart. My own round-3 mutation run had shown both surviving; I wrote
the category anyway. `--src-prefix` is a third.
Fixed by naming clauses instead of a category, because the category is where
the mis-sort hid. Then verified the sort rather than asserting it: all 19
clauses the sentence sorts were disarmed individually — the 12 it calls proven
were witnessed red, the 7 it calls unproven were witnessed green. The sentence
now matches a measurement.
Two stale universals removed with it. `docs/ci-cd.md` said "one honest gap"
where the boundary names seven; it now points at the boundary rather than
carrying its own count. And the FORMAT_KNOBS header comment still said "every
knob that reshapes the output this script parses" — the exact universal that
was blocked in `mechanics:` last round and simply moved one file over.
`diff.mnemonicPrefix` is named there as a live counter-example: it reshapes
the header, has no row, and is harmless only because --dst-prefix beats it.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Round 6 verdict was MERGEABLE with two follow-ups; both are one clause and
both are exactly this PR's subject, so they land here rather than in a
follow-up issue.
F1 I wrote that `diff.mnemonicPrefix` is "harmless only because --dst-prefix
beats it". The pin does beat it, but that is not why it is inert here.
Measured: git only uses mnemonic prefixes when a diff side is the worktree
or the index, and run_diff issues the three-dot `base...HEAD` form, where
git emits plain a/…b/ either way — with or without the pin.
git -c diff.mnemonicPrefix=true diff -U0 HEAD~1 -> +++ w/docs/f.md
git -c diff.mnemonicPrefix=true diff -U0 HEAD~1...HEAD -> +++ b/docs/f.md
A right observation attracting a wrong "because" is a shape this repo has
on record, and it is worse in a comment whose whole job is to stop a future
reader re-deriving something.
F2 the UNPROVEN list read as exhaustive while three clauses sat on neither
side. Each is now listed and each was disarmed to confirm the sort — the
`+++ /dev/null` arm, the malformed-`@@` arm and the stderr relay all
survive, matching the UNPROVEN claim. Both lists now say they ENUMERATE,
so a clause on neither has simply not been measured, rather than being
implied covered.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Round 7 verdict was MERGEABLE with one nit — an over-long comment line. Ruff
found two more while I was there: UP031 (percent format) and S110 (a bare
try/except/pass). The S110 is deliberate and stays: it guards the stdio
reconfigure, which IS the never-fails invariant's own setup and so is the one
place that must not raise. It is now noqa'd with that reason rather than left
looking like an oversight.
`ruff format` is deliberately NOT applied. This repo checks in no ruff config
and 21 of its 35 tracked python files fail `ruff format --check`, so
reformatting these two alone would diverge from every file around them — the
deviation policy says match the established style. `ruff check` is clean,
which is the half that finds defects.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
Closes #784.
Generalizes the no-session-narrative rule from
docs/handoffs/chicorytv-issue-queue.md(where it was correct but scoped to one file, for a file-specific reason) to every durable artifact, with the destination and the carve-out both stated.What ships
docs/decisions/records/docs/no-session-narrative.md(docs.no-session-narrative) — a doc records the END STATE; the path to it goes in the commit message, the issue, or the## Closing record. Concretely: a review finding is answered in the commit message and only the corrected claim enters the doc. Carries the who-benefits test and the carve-out (dated measurements, stated snapshot boundaries, tested-and-rejected negative results, the why behind a non-obvious choice, a trap and its consequence), plus the wholesale exemption fordocs/decisions/**.scripts/check-doc-narrative.sh+ adocs-reminderstep — ADVISORY. Exits 0 on every path; emits::warning::only.--diff <base>scans lines ADDED in the PR (the CI mode);--allsweeps the tracked corpus.docs/guard-inventory.md(TOOLING) anddocs/remote-state-inventory.md(N/A), plus adocs/README.mdtask-signal row.Why advisory and not a gate
A narrative detector is a string predicate over prose.
docs/defect-shapes-773.md§4 andtesting.guard-derives-population-from-sourceboth argue that class must not be load-bearing, and the withdrawntest_review_verdict_vocabulary_parity.py(six rounds, then deleted) is the empirical case. The record says in as many words that this may never become a blocking gate.Evidence it is a weak predicate, measured rather than asserted: on the same corpus a reading sweep found 18 sites and this detector found 10 — a union of 21, with each side missing sites the other caught.
Verification
scripts/tests/— 771 passed, 2 skipped.test_the_inventory_covers_exactly_the_guards_that_existandtest_every_in_scope_file_has_a_row_and_every_row_names_a_real_fileeach named the new script).--diff: a throwaway commit adding a narrative line todocs/testing.md, a matching line to a decision record, and one to the script itself → flagged exactly thedocs/testing.mdline, at the correct line number, exempted the record and the script, exit 0.nothing to flag, exit 0.decisions_validate.py,build_decisions_catalog.py --check,check-kickoff-guard.shall green.Deferred, stated rather than smuggled
The sweep reports and does not remediate. 21 existing sites across 6 docs are listed on #784; each needs the who-benefits judgement applied by someone who can see whether the prior-state note is load-bearing (several in
docs/remote-state-inventory.mdare — "graded down by cold review" tells the next reader not to regrade it back), and this repo has the inverse failure on record (#542, pruning the only copy). Follow-up filed.Cold review (BLOCKED) falsified three of the four things the PR asked it to attack. Reproduced each before acting: B1 the script could exit NON-ZERO on two reachable paths, against an invariant asserted absolutely in four places H1 an unresolvable base ref printed "nothing to flag" and exited 0, having scanned zero files — indistinguishable from a clean run M1 `'\\ No newline'*` never matched (two literal backslashes inside single quotes; git emits one), so the marker counted as content M2 an added line whose own text began `++ ` was eaten by the `+++ ` header arm. Verified against the real predecessor: it reported line 2 where the truth is 3. M3 core.quotePath hid non-ASCII paths from BOTH modes, silently M4 --all dropped a final unterminated line Those are one mistake at six sites: deciding what a diff line IS from its prefix, with no hunk state. Patching them individually is the shape that does not converge here, so the mechanism is replaced — check-doc-narrative.py parses with hunk state, uses core.quotePath=false, keeps rename detection, and returns 0 from a bare `except` as the last line of defence. H2 said TOOLING/NONE is the classification under which no proof is ever owed, and that a 99-line parser shipped with four demonstrated defects a fixture directory would have caught. Correct: scripts/tests/test_check_doc_narrative.py now asserts exit 0 per argument shape, compares every reported line number against truth read back off disk (never a hand-written number), covers each defect above, and carries a mutation proof that disarms PATTERNS so the suite cannot pass on a detector that finds nothing. M5/M6: docs/ci-cd.md now describes both steps of the job, and the record's mechanics: states the detector's population instead of implying it matches the rule's — skills and nested markdown outside docs/ are out of scope. Proves: scripts/tests/test_check_doc_narrative.py Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>Round 4 confirmed the git-format channel is genuinely closed — an independent 46-configuration matrix across two modes found zero false cleans, and all eight claimed mutant kills reproduced. What it blocked on was three sentences, in a PR whose whole subject is that a doc asserting a proof it lacks is what a later reader trusts instead of re-checking. Mine did it three times. M-C the workflow comment named "a checkout hiccup" as the motivating failure and then left Checkout unguarded, so its closing claim was false of the code it annotated. Checkout is now stated as deliberately OUT of scope: a job that cannot run is a different failure from an advisory one that ran and disagreed. The comment also records what I measured on the real runner (PR#811, run 2179) rather than what I assumed — both steps green under continue-on-error, job success, commit status success. M-F `# -U0 is what the line arithmetic assumes` is false: removing -U0 reddens nothing, because the context arm handles any -U. The row pins that a configured diff.context cannot beat the CLI, and now says so. N-E `mechanics:` claimed EVERY knob that can reshape the output has a row. Six knobs do not — they are harmless, which is why. Worse, the claim is unprovable in principle, which is the exact argument this fix used to close the channel instead of enumerating it. And the reviewer's larger point, taken: rounds 3 and 4 each converted "N load-bearing clauses have no mutant" into "N kills", and round 4 produced six more from the same well. That is two rounds of one mechanism, so the answer is to stop enumerating clauses, not to add five more tests. The record now carries the COVERAGE BOUNDARY in one sentence — the parse path and the format pins are mutation-proven; the error and degradation arms are defensive and unproven — and the two inventories point at it instead of repeating a flat "exits 0 on every path". Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>Round 6 verdict was MERGEABLE with two follow-ups; both are one clause and both are exactly this PR's subject, so they land here rather than in a follow-up issue. F1 I wrote that `diff.mnemonicPrefix` is "harmless only because --dst-prefix beats it". The pin does beat it, but that is not why it is inert here. Measured: git only uses mnemonic prefixes when a diff side is the worktree or the index, and run_diff issues the three-dot `base...HEAD` form, where git emits plain a/…b/ either way — with or without the pin. git -c diff.mnemonicPrefix=true diff -U0 HEAD~1 -> +++ w/docs/f.md git -c diff.mnemonicPrefix=true diff -U0 HEAD~1...HEAD -> +++ b/docs/f.md A right observation attracting a wrong "because" is a shape this repo has on record, and it is worse in a comment whose whole job is to stop a future reader re-deriving something. F2 the UNPROVEN list read as exhaustive while three clauses sat on neither side. Each is now listed and each was disarmed to confirm the sort — the `+++ /dev/null` arm, the malformed-`@@` arm and the stderr relay all survive, matching the UNPROVEN claim. Both lists now say they ENUMERATE, so a clause on neither has simply not been measured, rather than being implied covered. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>Review-verdict: MERGEABLE @
d938493Cold review, 8 rounds, final sha lint-only delta confirmed byte-identical. Channel closed: 46-config matrix, zero false cleans. Coverage boundary verified clause-by-clause by mutation, 19/19 sort correct. 807 passed, 2 skipped.