fix(bin): stop reading an unanswered backend probe as a dead endpoint

[RooseveltAdvisors]

Intent

Captain's order, 2026-09-05: "go close the last 2 gaps." Gap G7 (stale-claim reclaim) fails its automated reading only because bin/fm-crew-state.sh reports a task's endpoint as "backend target gone" whenever the herdr CLI errors or stalls under load, so a busy box briefly scores dozens of live claims as dead (76 at 02:4x EDT versus 9 five minutes later on an idle box). The sweep in bin/fm-stale-sweep.sh trusts that text as positive evidence of death. Make the reader tell the truth: a pane that is absent is gone; a CLI that failed to answer is unknown.

What Changed

• bin/fm-crew-state.sh: in the no-run fallback, a failed pane capture no longer unconditionally emits backend target gone. For the classifier-backed backends (tmux and herdr) it now consults fm_backend_agent_state and branches per verdict — missing/dead keep the gone-class text (the dead case naming it agent gone, pane shell remains), alive falls through to normal state classification instead of being discarded, and any other verdict emits backend unreachable (<backend> endpoint state: …). Backends without a classifier (orca, zellij, cmux) keep the previous capture-failure-means-gone reading.
• tests/fm-crew-state.test.sh: extends the fake tmux/herdr binaries with FM_FAKE_TMUX_UNREADABLE, FM_FAKE_HERDR_READ_FAIL, FM_FAKE_HERDR_HUSK, a list-windows inventory arm and a pane get arm, and adds four assertions/tests — a stalled herdr CLI and an unreadable tmux read unreachable rather than gone, an alive endpoint with a failed scrollback read still reports working, a husk pane still reports gone, and the existing dead-window test now asserts the gone text explicitly.
• tests/fm-backend-herdr.test.sh: pins the suite's ambient FM_HOME to a fresh marker-free directory so a secondmate-marked checkout can't flip the default workspace label for tests that don't set FM_HOME themselves.

Risk Assessment

✅ Low: The change is narrowly scoped to one fallback branch, preserves every positive-death path that licenses reclaim, routes both backends through an existing recovery-grade classifier rather than adding new state, is covered by four new behavior tests that exercise the real script, and leaves the non-classifier backends untouched exactly as the recorded decisions required.

Testing

Completed 1 recorded test check.

• Outcome: ⚠️ 1 error across 1 run (1m42s)

Pipeline

Updates from git push no-mistakes

✅ **intent** - passed

✅ No issues found.

✅ **Rebase** - passed

✅ No issues found.

⚠️ **Review** - 1 info

• ℹ️ bin/fm-crew-state.sh:570 - The fallback's section header states an unqualified invariant - "only positive evidence that the target is gone may read as death - a backend that failed to answer is unknown, never death, for both classifier-backed backends (tmux and herdr)" - but the block comment 20 lines below explicitly carves out the opposite for tmux: "tmux gave one of its definitive no-session/no-server/no-socket responses ... a socket-connection failure is NOT covered by the unknown-never-death rule above". Both are reachable: bin/backends/tmux.sh:312 maps error connecting to <socket> (No such file or directory) and (Connection refused) - a tmux that failed to answer - to missing, which this block emits as backend target gone. The same over-broad sentence appears in the file header at line 67 ("an endpoint that merely failed to answer reports unknown · none as unreachable"). This is comment accuracy only; the code behaves as the inner comment describes and as the round-2 decision ruled. Smallest remedy is to qualify both sentences to "a backend that failed to give a definitive answer", so the header no longer contradicts the exception it points at. No behavior change - making socket errors non-death would break fm-bootstrap.sh:793's respawn license and needs authorization, which is why only the wording should move.

⚠️ **Test** - 1 error

• 🚨 tests failed with exit code 1
• bin/fm-test-run.sh --changed --exclude-family real-herdr-gated

⚠️ **Document** - 1 info

• ℹ️ docs/fm-test-portable-shards.md:18 - The change adds four tests to tests/fm-crew-state.test.sh, whose 17558 ms entry in the parallel-lane timing table (and in docs/fm-test-isolation-proof.{md,json}) feeds the longest-processing-time lane assignment. I did not touch those numbers: they are explicitly dated measurements from the 2026-08-20 concurrent isolation proof (maintainer-verification), so hand-editing them would fabricate evidence, and regenerating them means running bin/fm-test-isolation-proof.sh, which is outside this documentation phase. The recorded facts are not false — they remain an accurate record of that run — but the derived lane-balance estimate is now slightly optimistic. Follow-up: refresh the proof and both tables the next time the isolated candidate set or its timings are re-measured.

✅ **Lint** - passed

✅ No issues found.

✅ **Push** - passed

✅ No issues found.

---

Note: the local test-gate failures for tests/fm-test-run.test.sh (a comm/LC_ALL sort artifact of the runner, zero failing assertions) and tests/fm-calm-pi-extension.test.sh (a parallel-load flake) are pre-existing and environmental - both reproduce identically on the untouched base commit and are unrelated to this change; upstream CI is green on all 15 checks.

[greptile-apps]

Confidence Score: 5/5

The PR appears safe to merge based on the reviewed changes, with no concrete blocking or independently actionable defect identified.

The classifier-backed fallback preserves positive death signals while converting unanswered probes to unknown and retaining normal classification for verified live endpoints; the accompanying tests cover the principal changed paths.

_{Reviews (1): Last reviewed commit: "no-mistakes(document): document classifi..." | Re-trigger Greptile}

[kunchenguid]

Speaking as Kun's firstmate: this is merged. Thank you @RooseveltAdvisors — really appreciate you taking the time on this.

VISION: (1) OK (2) OK (3) OK (4) restart/durability honesty restored — unanswered backend probe ≠ dead (5) OK (6) OK (7) scope OK — narrow fallback branch; positive death evidence only may reclaim.