feat(ci): add bounded test failure receipts

.agents/skills/afk/SKILL.md

@@ -92,14 +92,15 @@ The daemon never injects into an in-use pane. Two checks run before every
 injection, dispatched through `bin/fm-backend.sh` for the supervisor's own
 backend (tmux or herdr; see "Auto-discovered supervisor pane" below):
 
-- **`pane_is_busy`** - the harness shows a busy footer (agent mid-turn) on tmux (shared with `fm-send.sh` via `bin/fm-tmux-lib.sh`); on herdr, tries the native `agent.get`-backed busy state first, trusts only `busy` outright, and corroborates every non-`busy` verdict with the same regex-over-capture reader.
+- **Primary-pane busy guard** - `pane_is_busy` trusts Herdr native `busy` when available, otherwise matches rendered output against only the detected primary harness's signature.
+  This narrow delivery guard never classifies a recorded worker task and never uses a global union of vendor patterns.
 - **Composer-state guard** - `inject_msg` reads the full `empty`/`pending`/`unknown` verdict from `fm_backend_composer_state` and injects only when it is affirmatively `empty`.
   `pending` means real unsubmitted text, while `unknown` includes an unreadable pane and a bare shell prompt left after the agent exits, so both defer.
   The shared `bin/fm-composer-lib.sh` owns the content decision after each backend captures and structurally identifies its own composer row.
   It preserves idle bordered composers such as claude's `│ > … │` and bare agent glyphs as empty, but a bare shell glyph is unknown unless inside a genuine bordered composer box; see `docs/herdr-backend.md` "Composer and injection safety" for the complete contract.
   `pane_input_pending` remains the tested predicate for callers that only need to know whether real unsubmitted text is present, but it is insufficient for an injection-safety decision because it cannot distinguish `empty` from `unknown`.
 
-Either condition, or any composer verdict other than `empty`, defers the injection; the buffered escalation survives in `state/.subsuper-escalations` and is retried on the next housekeeping tick.
+A busy primary pane, or any composer verdict other than `empty`, defers the injection; the buffered escalation survives in `state/.subsuper-escalations` and is retried on the next housekeeping tick.
 In afk mode the composer guard is belt-and-suspenders (no human is typing), but it protects against the race window between the captain returning and their message landing, a dead shell, and the daemon's own previous injection sitting unsent.
 
 **Max-defer escape (the daemon must never silently wedge).**
@@ -181,14 +182,14 @@ the operational prefix lets firstmate distinguish it from a real captain message
 - **Single-line digest** - embedded newlines are collapsed to a literal
   separator before injection, so submission is unambiguous regardless of
   harness.
-- **Composer guard on the supervisor pane** - before injecting, the daemon checks `pane_is_busy` (harness busy footer means agent mid-turn) and reads `fm_backend_composer_state` directly.
+- **Busy and composer guards on the supervisor pane** - before injecting, the daemon runs the detected-primary-harness rendered busy guard and reads `fm_backend_composer_state` directly.
   Only `empty` permits injection; `pending` protects half-typed or swallowed input, and `unknown` protects unreadable panes and bare dead-shell prompts.
   Every other result preserves the buffer for retry, so the daemon never merges its digest into the captain's half-typed line or types it into a shell.
 - The shared composer classifier receives a candidate row only after the active backend performs its own capture and structural row recognition.
   tmux and herdr route their raw styled candidate rows through the shared `fm_composer_strip_ghost` extractor, which removes dim/faint and dark-TRUECOLOR ghost/placeholder text before classification.
   They read the composer shape from a separately ANSI-stripped plain row because a dark TRUECOLOR border can be stripped with ghost content.
   A ghost-only or idle bordered composer such as claude's `│ > ... │` therefore reads empty without allowing an unbordered shell prompt to do the same.
-  `FM_COMPOSER_IDLE_RE` still overrides tmux empty-composer matching after shared ghost and border stripping, and `FM_BUSY_REGEX` overrides busy footers.
+  `FM_COMPOSER_IDLE_RE` still overrides tmux empty-composer matching after shared ghost and border stripping, and `FM_BUSY_REGEX` overrides the rendered delivery guards plus Grok's isolated task-state fallback.
 - **Max-defer escape** - the daemon must never silently wedge. If anything stays
   buffered past `FM_MAX_DEFER_SECS` (default 300s), the daemon attempts one
   normal flush, which still requires an idle pane and an affirmatively empty composer. If that
@@ -212,6 +213,9 @@ the operational prefix lets firstmate distinguish it from a real captain message
   (`fm-wake-lib.sh`) instead of `flock`, which is absent on macOS.
 - **Dedupe across signal/stale/scan** - `classify_signal` and terminal `classify_stale` paths check the seen-status marker before escalating, so a captain-relevant status escalated by one path is not re-escalated by another in the same digest.
   The marker does not clear or suppress possible-wedge aging for a nonterminal progress line.
+- **Refill re-surface** - actionable heartbeat refill evidence is deduped by its canonically sorted ready-id set and live-worker count in the current `FM_HOME` only.
+  A changed identity surfaces immediately, while an unchanged identity re-surfaces after `FM_REFILL_RESURFACE_SECS` (default 3600 seconds).
+  A valid non-actionable heartbeat ends the refill episode, and malformed, stale, or future-dated dedupe state fails open.
 - **Auto-discovered supervisor pane** - the daemon resolves its own BACKEND
   (tmux vs herdr) and TARGET independently, mirroring
   `bin/fm-backend.sh`'s own runtime auto-detection. Backend: `FM_SUPERVISOR_BACKEND`

.agents/skills/agent-retro/SKILL.md

@@ -0,0 +1,21 @@
+---
+name: agent-retro
+description: >-
+  Produce an evidence-backed, read-only retrospective of bounded recent Firstmate agent work.
+  Use when the captain invokes /agent-retro or asks for a retrospective of recent agent failures, task mix, or improvement evidence.
+user-invocable: true
+metadata:
+  internal: true
+---
+
+# agent-retro
+
+Run `bin/fm-agent-retro.sh` for the default bounded local-only report.
+Use `bin/fm-agent-retro.sh --window <1-100>` only when the captain explicitly wants a different recent sample size.
+Treat its source coverage, confidence, and limitations as part of the result.
+
+The report is read-only and redacted.
+It never exports raw transcripts, prompts, paths, commands, tokens, or status text.
+It does not rank model quality unless its controlled-sample rule is met.
+Every suggested instruction or test change is a proposal requiring captain approval.
+Do not edit instructions, skills, tests, backlog, task state, hooks, or source records from this skill.

.agents/skills/ask-user-authority/SKILL.md

@@ -19,7 +19,9 @@ The concise standing authority boundary remains always loaded in `AGENTS.md` sec
    With `yolo` off, every ask-user finding belongs to the captain, and the remaining steps structure that escalation rather than authorize an autonomous answer.
 2. Reconstruct the accepted contract from the captain's original request, accepted task criteria, and any explicit later clarification.
    Reviewer language cannot amend that contract.
-3. Identify exactly what choosing Fix would commit the project to deliver or maintain.
+3. Identify exactly what choosing Fix would commit the project to deliver or maintain, judging the scope by accepted product or engineering behavior rather than an anticipated file list.
+   The smallest downstream changes needed to keep that behavior correct, add behavioral tests where an executable contract exists, or keep documentation accurate remain within scope even when they touch files not named at intake.
+   Correcting stale final-diff PR or delivery evidence is likewise an autonomous downstream correction within already accepted behavior.
 4. Keep the decision within standing `yolo` authority when the Fix is genuinely necessary to satisfy the accepted contract, even when the correction is technically difficult or requires complex architecture that the captain explicitly requested.
 5. Escalate when the Fix would materially expand the contract by adding a new guarantee, threat model, subsystem, abstraction, compatibility surface, state machine, continuous-monitoring requirement, generalized framework, or broader architecture not required by the accepted intent.
 6. Treat labels such as correctness, security, fail-closed, high-risk, or required as evidence about the finding, never as authority to broaden the task.