test: give every tracked evidence file exactly one writer

[FHoffarth]

Closes #3.

tracked evidence = deterministic + stable + single owner
run measurements = ephemeral + ignored (audit/run/)

The defect, reproduced rather than assumed

Two audit tests appended to a file another test owns and rewrites whole. Vitest runs
test files in parallel, so what the tracked file ended up holding depended on which of
the pair finished last. Forcing the adverse order with vitest's own file selection — no
test hack — destroyed recorded evidence, silently, with every test still green:

order run	evidence lost
gaps.test.ts → evidence.test.ts	4 lines
icc-check.test.ts → review-attacks.test.ts	2 lines

A full inventory of every writeFileSync / appendFileSync in audit/ and src/
confirmed these were the only two cases — the judgement is not based on filenames.

The fix

Both appenders record real evidence — findings, verdicts, canary survival — not timings,
so they get their own tracked file rather than the ignored run log.

evidence file	writer before	writer now
audit/evidence-notes.txt	evidence.test.ts + gaps.test.ts	evidence.test.ts
audit/review-attacks.txt	review-attacks.test.ts + icc-check.test.ts	review-attacks.test.ts
audit/gaps-notes.txt (new)	—	gaps.test.ts
audit/icc-notes.txt (new)	—	icc-check.test.ts

No citable evidence text changed. The five lines move across verbatim into the file of
the test that produced them. Nothing is reworded, reordered or dropped, and neither
long-standing file is restructured — the remaining diff lines are "no newline at end of
file" shifts.

Regression guard

EVIDENCE_OWNERS in the harness names the owner of each artifact;
audit/evidence-ownership.test.ts checks that table against what the sources actually
write. It counts every write call against the ones whose path it can read, so a path
assembled at runtime fails the check instead of slipping past it.

Verified by reintroducing both mistakes:

injected regression	guard
a second writer for review-attacks.txt	FAIL — "is written by more than one test file"
a write path built at runtime	FAIL — "writes to a path this check cannot read"

An earlier version of the guard asserted that each declared artifact exists. That check was
itself order-dependent — it ran before the new files had been produced — which is precisely
the fault this PR removes. It was dropped.

Evidence

Windows, core.autocrlf=true:

Gate	Result
Product tests	5 files / 43 tests passed
Audit tests	21 files / 296 tests passed (was 20 / 290)
tsc --noEmit	clean
vite build	204 modules, 4.26s
npm audit --audit-level=high	0 vulnerabilities
git diff --check	clean
git status --porcelain after a full audit run on the committed tree	empty
repeated full audit runs	clean both times
both previously destructive orderings, owner last	clean

npm ci could not run in the working checkout: Windows held an EPERM lock on
@rolldown/binding-win32-x64-msvc. Rather than kill unattributable node processes, the
lockfile was verified by running npm ci against a copy of package.json and
package-lock.json in a scratch directory — clean, 0 vulnerabilities. CI runs the real
npm ci on a fresh runner.

Merge gate

Full product and audit suites pass, every tracked evidence file has exactly one writer, and
a complete audit run leaves the tracked working tree clean on Windows with
core.autocrlf=true — including under the orderings that previously destroyed evidence.

Not in this PR

package-lock.json records "license": "ISC" while package.json says "MIT", so any
npm install rewrites the lockfile. Same class as this sprint, but out of scope; the change
was reverted and is left for a separate decision.

No product, parser or verification logic is touched. No test removed or weakened. No tag,
release or deploy proposed.

🤖 Generated with Claude Code