fix(hooks): make the dirty list mean "edited since last publish" (#4 + #5) - #11

Closes #4. Closes #5.

#5 carried two distinct problems. This PR fixes problem (1), the dirty-list. Problem (2) — basalt status reporting modified right after a successful publish for docs with local images — is a CLI-side comparator bug and is carried forward as #12 so that closing #5 does not silently close it too.

The proposition the hook got wrong

stop.sh asserted "this doc is unpublished" from a measurement of "this doc was edited". Those are different propositions, and two flows sat entirely inside the gap:

• P0-3: stop.sh nagea 'unpublished' sobre docs ya publicados por la Action del repo #4 — a repo-backed vault publishes via commit + push → Action. Nothing local ever cleared the list, so the nag was unconditional.
• El Stop hook pide publicar docs ya publicados (dirty-list sin publish-state) + status miente en docs con imágenes #5 — the agent runs basalt publish <file> mid-turn, which is exactly what the nudge asks for, and stop.sh was the only writer that ever cleared the list.

So the agent had no way to satisfy the hook except by not editing. A nudge that cannot be satisfied is not a nudge — it trains the reader to mute the channel, and then the sensor is worth less than nothing.

The list now means: vault docs edited since their last publish.

Two mechanisms, both local — no network added to the hot path

A — post-publish.sh (new), PostToolUse on Bash. Edit-tracking's twin: it appends a tombstone for a doc a command really published. The list became an append-only journal (post-edit.sh appends paths, post-publish.sh appends -<path>, stop.sh replays), so ordering is correct by construction and two O_APPEND writers cannot lose each other's updates.

B — a git-state filter in stop.sh. Nine conditions, all local. Its claim is deliberately narrow: positive local evidence that a push of this file, on this branch, starts a Basalt publish workflow. It does not claim to predict what the Action publishes — the real publish set is decided by a git ls-files | xargs pipeline and filename conventions in shell, which no hook can model. That limit is written into the code, not left implied.

Why not basalt status

It is the obvious fix and it is rejected on two independent grounds. It is network on every Stop with any dirty file, which #4's guardrail closes textually. And #5's own body documents it lying: publish rewrites image refs without touching the local file, so status returns modified immediately after a successful publish for any doc with local images — nearly every doc in the vault where the bug was filed. The fix would have failed exactly where it was reported, and failed silently.

Measured, not assumed

Three harness facts this rests on were captured, because none were previously verified — the existing capability notes cover Edit|Write; Bash had never been captured.

• PostToolUse(Bash) fires only on exit 0. Measured with a control counting executions separately from hook fires: 4 executions, 2 fires. A failed publish never reaches the hook, which is how success is known without an exit-code field (the payload has none) and without a network call.
• The payload's cwd does not follow a compound cd. cd /tmp && pwd prints /tmp; the payload still reports the session dir. Since El Stop hook pide publicar docs ya publicados (dirty-list sin publish-state) + status miente en docs con imágenes #5's own repro is cd <vault> && basalt publish …, resolving operands against cwd alone would match nothing — and a hook that matches nothing is indistinguishable from a dead one.
• ugrep and GNU grep disagree on [^\n] inside a bracket. Local dev and CI would diverge, so there is no regex path parsing here — case globs and string operators only.

Everything ambiguous fails toward the nag

--dry-run · publish f || true and non-final ; segments (exit 0 is a property of the command, not a segment) · quoted or multi-line commands · unresolvable operands · a doc whose extension the workflow's paths: list does not match · a workflow that is untracked, dirty, unpushed, has no push: trigger, uses paths-ignore: or ! negation, or is gated on another branch.

Verification

• Suite 83 pass / 0 fail (was 62). scripts/validate-skills.sh green.
• Mutation pass (tests/mutate.sh, opt-in, not CI): 15 mutants, 15 killed, 0 survived. Each mutant is proven live — patch lands, bash -n, smoke run — before its verdict is read, so "the test does not catch it" cannot be confused with "the mutant never ran". The first run killed 13 and two survived: one condition was shadowed by another, and one guard was defended in three places so removing a single layer changed nothing. Both are now isolated and killed.
• Positive control: the new suite run against the pre-fix hooks goes red — 29 failures, reproducing P0-3: stop.sh nagea 'unpublished' sobre docs ya publicados por la Action del repo #4 and El Stop hook pide publicar docs ya publicados (dirty-list sin publish-state) + status miente en docs con imágenes #5 verbatim. Critically, the negative arm behaves identically before and after: "edited and not committed" still nags. The feature was repaired, not deleted.

Explicitly not fixed here

• basalt status reports modified after a successful publish for docs with local images (El Stop hook pide publicar docs ya publicados (dirty-list sin publish-state) + status miente en docs con imágenes #5 problem 2) → basalt status reports modified right after a successful publish for docs with local images #12. The comparator lives in the CLI, not this repo.
• Publishing via the bundled MCP path is not tracked: its payload was never captured and publish_doc takes title + markdown rather than a path. A measurement gap, not a construction — and its failure direction is a surviving nag, never silence.
• The plugin treats .mdx as the primary intent-to-publish signal while the CI template the CLI emits publishes .md only → The plugin treats .mdx as the primary publish signal; the emitted CI template publishes .md only #13. This PR's hook is correct under both shapes because it evaluates the workflow's actual paths: list, but it does not resolve the product-level disagreement.