fix(bin): stop a resolve-key close from re-waking its home during signal grace

[jasonlgong]

Problem

• Part of The supervisor's own append to a task status file publishes a signal wake back to the supervisor #4767.
• When the watcher sees a changed signal, it waits one FM_SIGNAL_GRACE period, scans again, then classifies the first scan and the rescan together.
• If a fm-send --resolve-key answer is still committing its seen marker when the first scan runs, that scan records the answer's own status append as new.
• The rescan sees it as already handled, but the first-scan entry is still in the combined set.
• Result: the home wakes itself on its own answer.
• fix(bin): keep supervisor status closes from waking the same home #4895 and fix(bin): stop each keyed answer from re-waking this home #4907 fixed other paths to the same self-wake.
• This one still reproduces on current main (39f4c2a, with fix(bin): stop each keyed answer from re-waking this home #4907 merged): the new regression test fails there with the committed close survived the grace rescan.

Fix

• In bin/fm-watch.sh, after the grace period, status files are classified from the fresh rescan only.
• Turn-end markers from the first scan are kept, so a turn-end still surfaces if its marker disappears or changes back during grace.
• This relies on status logs being append-only, which is already the contract.
  • A status file rewritten in place at the same inode and size could hide a real signal.
  • That case is outside the contract and is documented in docs/architecture.md rather than handled here.

Tests

• tests/fm-send-resolve-key.test.sh: test_answer_close_during_signal_grace
  • Holds the answer mid-commit while the watcher scans.
  • Checks that the close does not wake the home, and that a later worker blocked: line still does.
  • Fails on 39f4c2a, passes with this change.
• tests/fm-watch-triage.test.sh: two cases cover a first-scan turn-end marker during grace.
  • One deletes the marker; one restores it to its reported signature.
  • Both still surface.

[kunchenguid]

Speaking as Kun's firstmate:

Tip 39d537c054d7bbdd89a884f9614c9ff166c7b8a8 — attestation MATCH. Contract-class restore (post-grace scan_signals only in bin/fm-watch.sh so a mid-grace --resolve-key seen-marker commit does not re-wake; append-only status contract documented).

Closes claim: author body explicitly says this does not fully close #4767 alone — it is the grace-window half. #4895 (fold/self-announce half) just landed as b6930db7. Keep this PR for the remaining grace path; do not treat GitHub closesReadyForPr as a full close vote.

Fork workflows approved this pass (CI 35448145592, NM 35448145611) — now in progress. MERGEABLE/UNSTABLE until green. We will re-watch after CI + no-mistakes SUCCESS. Waiting on CI, not on the author.