fix(wiki,knowledge-flush): enforceable prohibition rule, disjoint routing scopes, flush concurrency (#36 #37 #77)

[choiyounggi]

Closes #36
Closes #37
Closes #77

Three independent issues, implemented in parallel worktrees and merged locally
after review. 26 files, +1142 / -60.

#36 — wiki-lint check 2 is now mechanically enforceable

Check 2 flagged 147 prohibition phrases with zero true violations
(plans/wiki-audit/findings/structure.md), so it was permanent manual-review
noise that hid nothing.

The rule was validated against the live corpus over four iterations before any
code was written. Earlier formulations were rejected by measurement, not taste:

formulation	violations reported	why it was wrong
per line	138	descriptive never ("references never released")
per sentence	11	wrapped list items were split mid-item
per item	6	## Sources citations quote prohibitions
per item, scoped	0	ships

Final rule: a prohibition clause (≥3 words, starting with
don't/do not/never/avoid/must not) must share its directive item —
a table cell or a bullet — with either the replacement action or the
mechanism that makes it true. ## Sources is out of scope.

Corpus practice pairs prohibitions with a reason more often than with an
alternative action (Never retry — the request itself is wrong; the same bytes fail again), which the old AGENTS.md rule 3 would have failed. The new rule
permits it.

• scripts/wiki-lint-prohibitions.js — exits 0 clean / 1 violations / 2 bad dir
• Negative-control fixtures + 10 bats tests, so the checker is observed both
  passing and failing rather than only ever returning 0
• 2-word cells (| Unknown key | Never read | vs | Timeout | Never retry |)
  are undecidable by shape, so they surface as new check 11 at info

#37 — disjoint routing scopes for two ambiguous pairs

Routing probes 6 and 7 (plans/wiki-audit/findings/routing.md) were AMBIGUOUS:
two concerns each claimed verbatim by two domains.

• Doc gates split on mechanism vs acceptance: testing/quality owns
  authoring and validating the check; qa/document-verification owns whether
  passing it is enough to accept the deliverable.
• Flaky tests split on what is the unreliable thing: a test in your suite →
  testing/flaky; the system itself → debugging/concurrency.

The flaky predicate was first written as "do you have an on-demand
reproduction?" and the independent auditor rejected it twice — correctly. Probe
7's wording ("an intermittent CI test failure that passes on retry") does not
state whether a reproduction exists, so that predicate is undecidable at the
routing gate, which is exactly where it must be decided. The subject of the
failure is always stated in the symptom, so subject is the decidable axis.

Both probes re-run UNIQUE from index lines alone. Also fixes the INDEX.md
backend LLM phrasing (a weak attractor for ML-training queries) and adds a
databases→backup cross-pointer.

#77 — knowledge-flush no longer races itself

The manual skill and the auto-flush.sh hook drained one queue into competing
PRs and reset each other's checkout. Fixed on both axes the issue identified as
complementary — and the bundled wiki independently prescribes the same pairing
(distributed-locks: "the lock alone cannot provide this";
scheduled-job-overlap: "prevention lowers the probability; idempotency removes
the damage"):

• scripts/flush-lock.sh — one owner-token lock shared by both entry points.
  mkdir is the primitive because flock(1) is util-linux-only and absent on
  macOS. Release is delete-if-token-matches; reclaim needs TTL expiry and a
  dead pid.
• hooks/queue-claim.js — per-run claim of queue rows, so even a concurrent pair
  produces disjoint PRs. Atomic tmp+rename; an unparseable line is carried
  through verbatim, never dropped.

This also removes a live bug: auto-flush.sh released its old lock with a bare
rmdir, the exact anti-pattern that deletes the next holder's lock after a TTL
expiry.

Two blockers found in review and fixed before merge, both from the run id being
computed per process:

1. the hook acquired the lock and the headless session it spawned then hit step 0
   and got exit 3 — silently killing the whole auto-flush path. acquire is now
   re-entrant for the same owner; a foreign run id still gets 3.
2. the skill's step 0 and step 5 run in separate shells, so an unprefixed
   release computed a new run id, was refused, and leaked the lock for the full
   900s TTL.

Verification

• 485 bats tests, 0 failures on the merged tree, run in an isolated worktree
• wiki-audit: wiki-lint check 2 vs corpus — prohibition-with-replacement rows need a mechanically enforceable rule #36's checker over wiki-audit: routing ambiguity — doc-gate and flaky-test concerns have dual domain ownership #37's merged wiki: 61 directives, 0 violations — the two
  wiki-touching tasks do not break each other
• flush-lock end-to-end re-checked on the merged tree: re-entrant acquire 0,
  foreign acquire 3, cross-shell release 0, no leak

Every verification command was re-run by the coordinator rather than taken from
the worker reports.

Not included

#38 (wiki category seeds) and #75 (orchestrator-skill pattern) were in the
approved split but not implemented; #38 collides with in-flight work on the same
domain indexes. Both stay open.