fix(daemon): stop away-mode busy-guard false positives and refuse silent wedge-alarm entry

[ktapa]

Intent

Away-mode supervision on the claude harness with native-hosted injection was delivering zero escalations: the busy guard falsely reported the supervisor pane busy continuously (not transiently) even while idle, for over 2 hours across two separate sessions, leaving a task parked awaiting a decision with firstmate unaware. Establish by direct measurement (not by reasoning from source alone) whether the daemon's own tracked background-job hosting under start-native is what trips the pane busy-guard on claude, or whether the true cause lies elsewhere (a claude busy-signature matching an idle rendered pane, or a mis-resolved supervisor target), and state the finding plainly either way. Fix the root cause so an escalation raised while away mode is active reliably reaches its target pane on claude with herdr, demonstrated end to end in an isolated Herdr lab session (not just a passing unit test). Add regression coverage for the busy-guard false positive. Separately, fix the wedge alarm's fail-silent failure mode: on Linux there is no OS-level alert channel, so a busy-guard false positive can wedge silently with only an inert durable marker as the signal; either provide a working default alert channel on Linux or make the absence of one loud at away-mode entry (before the captain walks away) rather than only at failure time - prefer entry-time refusal when no reliable Linux channel is available. Check whether the wedge-alarm entry preflight (wedge_alarm_reliable_channel_configured) can itself be tricked into reporting a channel reliable when it cannot actually fire, and fix any such gap: specifically, a bare 'command:' directive with no payload was matching the command:* pattern and passing as reliable while the runtime command dispatch's own non-empty-payload guard would silently no-op it - fixed so an empty payload is rejected at entry like any other malformed directive, with regression coverage pinning both the bare-command rejection and the still-accepted command: case. Also determine whether the terminal-backed away-mode launch path (not just the native-hosted one) is affected by the busy-guard defect, and if only native hosting is broken, fix the afk skill's harness routing accordingly as part of this task.

What Changed

• bin/fm-supervise-daemon.sh's pane_is_busy now short-circuits to "not busy" whenever fm_backend_composer_state reads an affirmatively empty composer, before falling back to the unanchored rendered-tail scan — fixing a false-positive where settled scrollback text matching Claude's busy-shape signature (e.g. "…(2h into the soak test)") kept the pane marked busy indefinitely on both the terminal-backed (start) and native-hosted (start-native) launch paths, since both share the same pane_is_busy call.
• Extracted wedge-alarm channel resolution (wedge_alarm_configured_channels, wedge_alarm_platform_default) out of fm-supervise-daemon.sh into a new shared bin/fm-wedge-alarm-lib.sh, and added wedge_alarm_reliable_channel_configured, which rejects unrecognized/malformed directives and a payload-less bare command: directive that would otherwise pass as "reliable" while the runtime dispatch's non-empty-payload guard silently no-ops it.
• bin/fm-afk-launch.sh's start and start-native now call a new fm_afk_launch_wedge_alarm_preflight that refuses a fresh away-mode entry when no channel resolves to something that can actually fire (an explicit off still passes as a deliberate acknowledgment), so the missing-alert-channel gap is surfaced loudly before the captain walks away instead of only at failure time.
• Added regression coverage for the settled-scrollback busy-guard false positive (tests/fm-daemon.test.sh) and for the wedge-alarm entry preflight, including the bare-command: rejection and accepted command:<cmd> cases (tests/fm-afk-launch.test.sh), plus supporting docs updates (docs/wedge-alarm.md, docs/verification/supervision.md, docs/scripts.md, .agents/skills/afk/SKILL.md).

Risk Assessment

✅ Low: The branch cleanly implements the fix (composer-state short-circuit in pane_is_busy backed by pre-existing, well-tested composer classification, verified live per the docs), correctly resolves all three prior-round findings (vacuous rollback test now reaches the real failure path, the unvalidated-directive gap is closed via an explicit recognized-directive allowlist, and the bare command: payload is now rejected with command:?*), and adds direct unit coverage for the new preflight function; source, tests, and docs are consistent with each other and with the required intent (busy-guard root cause investigated and fixed, wedge-alarm entry-time refusal added, both launch paths confirmed to share the fix).

Testing

Targeted tests for this final commit's fix (bare command: wedge-alarm directive rejection) and the related daemon busy-guard suite both pass; a before/after revert check proved the new unit test is a genuine, non-vacuous regression test that fails without the fix and passes with it. No findings.

Evidence: wedge-alarm preflight regression test results

ok - wedge-alarm preflight: refuses a fresh entry with no configured channel and no platform default
ok - wedge-alarm preflight: an explicit 'off' acknowledgment passes entry
ok - wedge-alarm preflight: refuses entry on an unrecognized/malformed channel directive
ok - wedge-alarm preflight: refuses entry on a bare 'command:' directive with no payload
ok - wedge-alarm preflight: a 'command:<cmd>' directive with a real payload passes entry

Evidence: reverting the fix reproduces the bug (pre-fix failure)

not ok - wedge-alarm preflight: a bare 'command:' directive was accepted as reliable (rc=0):

Pipeline

Updates from git push no-mistakes

✅ **intent** - passed

✅ No issues found.

✅ **Rebase** - passed

✅ No issues found.

✅ **Review** - passed

✅ No issues found.

✅ **Test** - passed

✅ No issues found.

• bin/fm-test-run.sh tests/fm-afk-launch.test.sh tests/fm-daemon.test.sh
• Regression proof: reverted bin/fm-wedge-alarm-lib.sh's command:?* back to command:*, reran tests/fm-afk-launch.test.sh, confirmed unit_wedge_alarm_preflight_refuses_empty_command_directive failed (rc=0, bare command: wrongly accepted) — then restored the fix and reran to confirm all pass
• bin/fm-test-run.sh tests/fm-afk-launch.test.sh (post-restore full rerun, exit=0)
• git status --porcelain / git diff --stat (confirmed clean working tree, no transient artifacts left behind)

✅ **Document** - passed

✅ No issues found.

✅ **Lint** - passed

✅ No issues found.

✅ **Push** - passed

✅ No issues found.

[greptile-apps]

Confidence Score: 5/5

The PR appears safe to merge.

No blocking failure remains.

_{Reviews (3): Last reviewed commit: "no-mistakes(ci): Fixed the Greptile-flag..." | Re-trigger Greptile}