Watcher wedge-escalates a lane that is correctly parked at a validation gate, indefinitely

[AlexAgenCheck]

What happens

A lane parked at a no-mistakes validation gate is correctly waiting on a human decision, but the watcher reads its silence as a wedge and escalates. The escalation repeats indefinitely, and each one costs the supervising firstmate a full turn.

Observed counts on a single home: one correctly-parked lane reached 671 consecutive wedge escalations, and another reached 70 before it was noticed.

Why the existing mitigations do not cover it

• A declared paused: status does nothing for a live ordinary crewmate. pause_state_class falls through to none for a non-secondmate live agent, so the declaration is inert exactly where it is needed.
• bin/fm-control.sh <id> exit on a finished lane does not reduce alerts either: the watcher treats a stopped agent like an idle one.
• Upstream b57c4d6 fixed the BUSY-pane sibling case only. The parked-at-gate case is still live as of 4179c50.

Current workaround

FM_STALE_ESCALATE_SECS and FM_PAUSE_RESURFACE_SECS are read by both the watcher and the away daemon, so raising them widens the window. This is confirmed in the daemon's stale_escalate= startup log. It is a blunt instrument: it delays genuine wedge detection by exactly the same amount.

Suggested direction

The watcher already has enough information to tell the two cases apart: a lane whose current run-step is a parked gate awaiting a human decision is not silent-because-wedged, it is silent-because-waiting. Treating a verified parked-at-gate state as a declared wait, on a long recheck cadence rather than a wedge escalation, would remove the noise without weakening real wedge detection.

Related: bin/fm-captain-hold.sh complete can write a captain-held [key=...] status line that the classifier does treat as a declared wait, which works well when it succeeds. That path is currently blocked in one common case, filed separately.

Environment

firstmate at 4179c50, tmux backend, claude harness.

[kunchenguid]

Speaking as Kun's firstmate:

Verified against main 3e5577b2c00fc90b1c79bc75e7281583f8553d7c.

VISION: aligns — Honest interface under load; never invent a wedge when the lane is correctly waiting; scripts own exact mechanics; unbounded false escalations burn captain/firstmate attention and tokens.

Main inspection: bin/fm-crew-state.sh correctly emits state: parked for awaiting_approval/fix_review gates. crew_absorb_class in bin/fm-classify-lib.sh explicitly maps parked to none (only working/paused absorb). pause_state_class in bin/fm-watch.sh returns none for a live non-secondmate even when status declares paused (agent_alive != dead). Result: a correctly parked validation-gate lane takes the wedge timer and escalates indefinitely. Defect still on main.

Related open PRs, not the precise help:

• patch(fm-watch): cap wedge escalations to prevent unattended LLM loop drain #2605 caps wedge escalations after N (bounds damage; does not treat parked-at-gate as a declared wait).
• fix(bin): absorb a decision-parked crew on the long re-surface cadence instead of wedge-escalating #2496 absorbs open-captain-decision + provably-working panes on the long cadence (different predicate than run-step parked).

Decision: ready-for-pr. Implement treating verified run-step parked as a declared wait on the long recheck cadence, without weakening real wedge detection. Do not open a competing PR against #2605/#2496 for their narrower scopes — this issue needs the parked-state absorb path.