feat(hooks): publish_overrides — a per-PATH policy that overrides the vault's (#10)

[A-PachecoT]

The hole

publish: in vault.yaml is per VAULT. The state that produces the Stop nudge is per
DOC. So when one doc cannot publish — because the server rejects it, not because
nobody tried — the only lever was publish: manual, which silences every other doc in
the vault too. What is left is a nag that fires every session with no action available,
which is how a channel gets trained into being ignored.

This repo already decided that case in the orphan bucket (stop.sh, "clear either way →
no re-nag"). This is the missing other half.

The shape

vault.yaml gains an optional block. Absent ⇒ today's behavior, byte for byte.

name: atelier
publish: prompt
publish_overrides:
  manual:
    - BITACORA.mdx                  # why this path is listed goes right here
    - docs/context-architecture.mdx

Two properties are the point: it is explicit (someone had to type the path, and by
convention the reason beside it) and it expires by itself (delete the line when the
block lifts). No implicit suppression, ever.

Every narrowing fails toward keeping the nag:

shape	behavior
exact path, relative to the vault root	overrides the vault policy
a glob (* ? [ ] { } !)	matches nothing — one line must not silence a subtree nobody enumerated
unrecognized key under publish_overrides:	ignored; the doc keeps the vault policy
flow style {manual: [a]}	reads as no overrides at all
unreadable vault, doc outside the root	delegates to vault_policy()

The vault root is the directory holding vault.yaml — the same root the CLI walks up to
when it derives a slug. One rule in both places or they drift.

Structure

vault_policy() is untouched. doc_policy() sits beside it, resolves an override and
otherwise delegates — so non-regression is structural rather than a property we have to
keep testing for. stop.sh changes on one line (the bucketing) plus the comment above it.

⚠️ auto is an accepted key, so a listed path can newly reach the bucket that shells out
to basalt publish. That bucket's logic is unchanged; only which files can reach it, and
only when someone wrote the path explicitly. It has its own test, plus the twin proving an
unlisted path in the same vault is never auto-published.

Measurement

bash plugins/basalt/hooks/tests/run.sh → rc 0, pass 83 → 100.

Every arm is a pair: the listed path that goes silent, and a path in the same session
that must still nag. The sharpest one is #10 (b) — override lists BITACORA.mdx, session
edits it and README.mdx, must nag for exactly 1. 0 means the vault got silenced
(feature deleted); 2 means the override did nothing.

Falsifiable-instrument gate — both mutants verified by effect before their verdicts
were read:

mutant	effect probe	result
force doc_policy to delegate always	listed=prompt (override dead)	arm (a) RED — the listed doc nags; 9 arms fail
force doc_policy to return manual	unlisted=manual (silence unconditional)	arm (b) RED — <silence> where exactly 1 was required; 35 arms fail

Also covered: all three vault policies with no publish_overrides key (unchanged),
./x.mdx normalization, a trailing # reason comment, block termination at column 0, and
both halves of the suffix trap — a bare x.mdx does not match notes/x.mdx, and an
identically-named doc in a different vault is not silenced.

bash scripts/validate-skills.sh → rc 0. The vault.yaml example in
skills/basalt/references/cli-path.md documents the new key.

Closes #10