fix: preserve watcher check wakes

[tommy230]

What

Reworked replacement for #52, rebased onto current main.

This makes watcher PR/check wakes lossless by moving repeated-output suppression into fm-watch.sh: check output is written to the durable wake queue before the watcher advances its .seen-check-* marker.

Why

state/*.check.sh was the remaining wake source whose suppression could happen inside an opaque check script. A legacy edge-triggered check could advance its own .babysit-*.seen marker before its stdout became a delivered watcher wake; if that stdout was lost by a timeout, crash, or concurrent watcher run, the terminal transition could be swallowed permanently.

That violates the durable wake invariant used by status and stale-pane signals: enqueue first, suppress second.

Changes

• Adds watcher-owned dedupe for stateless check output via .seen-check-*.
• Enqueues check wakes before advancing the dedupe marker.
• Keeps a catch-all for legacy .babysit-*.seen sidecars that already self-suppressed a terminal MERGED or CLOSED state.
• Marks matching legacy sidecars as already escalated when a normal check wake is delivered, avoiding duplicate catch-all wakes.
• Cleans check-related watcher state during teardown.
• Updates the PR-check contract docs and adds focused regression coverage.

Validation

• tests/fm-wake-queue.test.sh - passed
• tests/fm-teardown.test.sh - passed
• for test_script in tests/*.test.sh; do "$test_script"; done - passed
• bash -n bin/*.sh tests/*.sh - passed
• shellcheck -x bin/*.sh tests/*.sh - passed
• no-mistakes run 01KVWAJEV4K1D91FS6J8Q400ZX - checks-passed

The no-mistakes review found and fixed one issue in this branch before final validation: the legacy catch-all could duplicate a check wake after a normal check output had already delivered. The final branch includes that fix.

GitHub currently reports no status checks configured for this fork PR, so the validation evidence is the local/no-mistakes run above.