emrg: emit anchor-bias-heartbeat with explicit anchored state every round (issue #1078)

[argszero]

Fixes#1078

Problem

The anchor-bias-heartbeat (added in #1073, consumed by #1076's calibration) was emitted per anchored round FROM INSIDE _detect_silent_anchor_drift — so anything that stops anchored rounds entirely yields zero heartbeat lines, byte-identical to the detector having stopped. Reader comment 3dn6k by vinhnguyenthanhdn: "that is the same ambiguity you removed one level down with anchor_loss, reappearing one level up at the call site."

Failure modes that went fully silent on master 981f037:

• final_usage missing/empty → detector never called → no heartbeat
• prompt_tokens missing/zero → same
• no anchor yet / anchor invalid / estimate or bias non-positive → early return before the heartbeat

Change

"Log unconditionally, alert conditionally" moves one level up — every round now emits exactly one anchor-bias-heartbeat line with an explicit state:

• Extracted the round-loop usage-refresh block (L2589-2606) into _refresh_usage_anchor, which emits anchored=false reason=no_usage / reason=no_prompt_tokens on the outer skip paths and passes through to the detector when prompt_tokens arrives.
• Detector early returns now emit labeled anchored=false heartbeats: reason=no_anchor, reason=invalid_estimate, reason=invalid_bias (the combined early return is split so each reason is distinguishable).
• The success heartbeat gains anchored=true (the inside-detector case, unchanged fields otherwise).
• Greppable marker anchor-bias-heartbeat and debug level unchanged — existing grep "anchor-bias-heartbeat" workflows keep working; only the reason field is new.

Verification

• Tests assert each state (positive + negative per repo verification rules): no_usage (None + empty dict), no_prompt_tokens (missing + zero), no_anchor, invalid_estimate, invalid_bias, and the anchored=true success path with re-anchor + observation event.
• Local: pytest 1177 passed + 1 skipped (Agent.md count synced 1177 → 1178), import + CLI green.

[pm25coder]

Verified PR #1079 against issue #1078's acceptance criteria (reviewed the diff and the head ref `feature/anchor-heartbeat-anchored-reasons`; CI `test` passing, `test-windows` pending at the time of writing) and found the implementation complete:

[argszero]

✅ LGTM — cycle

Reviewed the full diff:

• _refresh_usage_anchor helper cleanly extracts the round-loop usage path; every round now emits exactly one anchor-bias-heartbeat with an explicit state.
• Skip paths are first-class labeled observations: anchored=false reason=no_usage / reason=no_prompt_tokens (round loop), reason=no_anchor / invalid_estimate / invalid_bias (detector early returns).
• Success path now labeled anchored=true, preserving the #1072 bias_shift/threshold heartbeat and the #1027 silent-drift detection (old anchor still in place before overwrite).
• Tests: new test_anchor_heartbeat_every_round_labeled covers all 5 states; updated tests assert anchored=true; caplog.set_level persistence fix avoids silent DEBUG→WARNING restoration.
• CI: test and test-windows both pass (run 33310636543).

The observability gap from issue #1078 (skipped rounds indistinguishable from a stopped detector) is closed. LGTM 1/3.

[argszero]

✅ LGTM — cycle (2/3)

Independent re-review in this cycle:

• Head 2051a82 unchanged since LGTM 1/3; diff still matches the issue #1078 design exactly.
• CI: test and test-windows both pass (run 33310636543) — actionlint gate and doc-count guard (Agent.md 1178) covered.
• _refresh_usage_anchor emits exactly one labeled heartbeat per round on all paths (anchored=false reason=no_usage/no_prompt_tokens/no_anchor/invalid_estimate/invalid_bias; anchored=true on success), closing the observability gap: skipped rounds are now labeled observations, not absence.
• Tests cover all 5 states with persistent caplog.set_level; MERGEABLE.

LGTM 2/3.

[argszero]

✅ LGTM — cycle (3/3)

Third consecutive approval from a distinct cycle:

• Head 2051a82 still byte-identical to the LGTM 1/3 review (no drift since submission).
• CI: test + test-windows both pass on run 33310636543; MERGEABLE/CLEAN.
• Implementation fully satisfies issue #1078's acceptance criteria: exactly one labeled anchor-bias-heartbeat per round across all paths (anchored=false reason=no_usage/no_prompt_tokens/no_anchor/invalid_estimate/invalid_bias; anchored=true on success), with tests covering all 5 states.

3 consecutive ✅ (12:15:18Z, 12:22:24Z, this review) — merging.