fix(afk): force-deliver wedged escalations past max-defer

[rfbr]

Problem

In away-mode (/afk), the sub-supervisor daemon detected a crewmate stall within minutes and buffered an escalation, but could not deliver it to firstmate's pane for ~8h — so the stall was never recovered and the whole away period was wasted (state/.subsuper-inject-wedged read WEDGED: 29325s undelivered).

Root cause

The max-defer escape in housekeeping() (block 1b) only re-tried the same guarded escalate_flush. It never force-delivered. So once the composer guard started deferring, delivery deferred forever:

1. A flush types the marker-prefixed digest, then the Enter is swallowed → the daemon's own unsubmitted digest is left sitting in the supervisor composer.
2. On every subsequent tick, pane_input_pending reads that stale self-injected text as "pending input" → the composer guard defers.
3. Block 1b re-tries the same guarded flush → hits the same guard → defers again, and only ever calls inject_wedge_alarm.
4. The alarm is invisible in afk — nobody is watching firstmate's pane.

Net: the daemon self-poisons its own input channel and never escapes.

Fix

Past FM_MAX_DEFER_SECS, on a pane that is not genuinely busy, the daemon now force-delivers: it clears the stale self-injected composer text and submits the digest.

• Genuinely busy mid-turn pane (agent active, e.g. esc to interrupt) → still defers and raises the visible wedge alarm; an active turn is never clobbered.
• Idle pane with stale/self-injected pending text → cleared + force-submitted.
• Force path is afk-only (block 1b is afk-gated), so when afk is OFF a human's half-typed line is never clobbered — the captain-return-race guard is fully preserved.
• Clear-before-type also prevents two sentinel-prefixed digests from concatenating into one corrupted turn (the self-poison is fixed at the source on this path).

Changes

• bin/fm-tmux-lib.sh — new fm_tmux_clear_composer (sends Ctrl-A/K/U, verifies empty, fails open on an unreadable pane so the force path never re-wedges).
• bin/fm-supervise-daemon.sh — thread a force flag through escalate_flush → inject_msg; block 1b force-flushes; busy-pane and afk-off behavior unchanged.
• .agents/skills/afk/SKILL.md — document the force-deliver contract.
• tests/fm-afk-wedge.test.sh — regression suite: force-delivers stale self-injection, busy pane still defers, afk-off never clobbers a human line, and the clear-composer primitive.
• tests/fm-daemon.test.sh / tests/wake-helpers.sh — the prior "max-defer on a pending composer alarms without typing" case encoded the old behavior; retargeted to a genuinely busy pane (the case that still defers) and taught the fake tmux to honor clearing keys.

Testing

This clone has no no-mistakes gate, so verification was done directly:

• bash -n bin/*.sh — clean.
• shellcheck bin/*.sh tests/*.sh — clean (shellcheck 0.10.0).
• New tests/fm-afk-wedge.test.sh (4 cases) — pass.
• Full behavior suite — all pass except three pre-existing, environment-only failures unrelated to this change (fm-bootstrap/fm-x-mode need jq; fm-teardown needs a git fixture), each confirmed failing identically on the clean base.

All tests are hermetic (temp dirs + fake tmux); none touch the live state/ of a running firstmate.