Emit the anchor-bias heartbeat from the round loop with an explicit anchored:false reason #1078

Emit the anchor-bias heartbeat from the round loop with an explicit anchored: false reason, so a skipped round is distinguishable from a detector that never saw it

Source

Reader comment 3dn6k by vinhnguyenthanhdn on "A guard that has never fired and a guard that stopped running look identical on disk" (https://dev.to/pm25coder/a-guard-that-has-never-fired-and-a-guard-that-stopped-running-look-identical-on-disk-177f), 2026-08-30T05:53Z.

Reader's insight

The anchor-bias-heartbeat (added in #1073, consumed by #1076's calibration) is 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:

"that is the same ambiguity you removed one level down with anchor_loss, reappearing one level up at the call site."

Current state (master 981f037)

• Heartbeat logged at emrg/server/daemon.py:3128-3135, inside _detect_silent_anchor_drift (L3094) — AFTER the early returns at L3114 (if not old or old[1] <= 0 or estimate <= 0: return) and L3119 (if old_bias <= 0: return).
• Call site at L2594-2606: if final_usage: → pt = final_usage.get("prompt_tokens") → if pt: → _detect_silent_anchor_drift(session, pt, estimate). The detector is only reached when the LLM response carries prompt_tokens AND a previous anchor exists.
• Failure modes that go fully silent:
  • final_usage missing/empty (config flag, upstream change, stream edge case) → detector never called → no heartbeat;
  • prompt_tokens missing/zero → same;
  • no anchor yet / anchor invalid (first round, post-compact, post-switch, provider stopped returning usage) → early return before the heartbeat.
• _warn_missing_usage_anchor (L2998-3050) covers only the auto-compact projection gate, fires once per session, and does not emit the drift heartbeat.

Proposed change

Move the heartbeat to the round loop so every round emits exactly one anchor-bias-heartbeat line with an explicit state:

1. Keep the existing inside-detector heartbeat as the anchored: true case (it already logs bias_shift / threshold / old_bias / new_bias).
2. In the round loop at the usage-processing point (L2594-2606), log anchored: false with a reason when the detector is skipped:
   • reason=no_usage — final_usage missing/empty;
   • reason=no_prompt_tokens — final_usage present but prompt_tokens missing/zero;
   • reason=no_anchor — no previous anchor (not old or old[1] <= 0);
   • reason=invalid_estimate — estimate <= 0;
   • reason=invalid_bias — old_bias <= 0 (kept as-is, reachable only via the detector's early return).
3. Keep the greppable marker anchor-bias-heartbeat and debug level unchanged so existing grep "anchor-bias-heartbeat" workflows keep working; only the reason field is new.

This preserves "log unconditionally, alert conditionally" one level higher: a skipped round is now a first-class, labeled observation instead of an absence.

Acceptance (self-verifiable)

• A round with no final_usage emits anchor-bias-heartbeat ... anchored=false reason=no_usage.
• A round with final_usage but no prompt_tokens emits anchored=false reason=no_prompt_tokens.
• A round with no previous anchor emits anchored=false reason=no_anchor.
• A normal anchored round still emits anchored=true with bias_shift (existing behavior unchanged).
• Tests assert each state (positive + negative per the repo's verification rules); CI test + test-windows green.