CDD: reconcile coherence from telemetry; analyze the response after a failed call

[b-macker]

Summary

Checking the repo-sentinel F-008 handoff (the raw telemetry for both Stage 4 runs) turned up two CDD defects and one repo-hygiene fix.

1. Coherence moved without any telemetry

F-008 reported that the listed penalties never matched the coherence drops. CDD's arithmetic was correct: each drop was exactly the penalties minus natural healing (0.03 / (1 + signals_fired)), to four decimals on all six turns. The healing just wasn't reported anywhere. Three other movements weren't reported either:

• temporal decay
• the clamp at 0, which discards any penalty beyond 0
• recoverCoherence(), called on a passed step-up challenge or a failed pipeline stage

CDD_TURN now carries a coherence_adjustments field, so every analyzed row can be reconciled from telemetry alone:

coherence = previous analyzed coherence - penalties + validation_recovery + sum(coherence_adjustments)

• Contents: temporal_decay=-, natural_healing=+, floor_absorbed=+, recovery=+.
• Decay and recovery happen between analyzed turns. They build up in DriftState.pending_* and are reported on the next analyzed row.
• It's a separate field on purpose. living-script_v3/report.py, its gate registry and test_signal_contract.sh all read a non-empty penalties_detail as "a signal paid this turn", and healing lands on nearly every turn after damage.
• validation_recovery now reports the credit actually received. It's capped at 1.0, so it can be less than the configured amount.

2. The response after a failed API call was never analyzed

Writing the reconciliation test exposed this one.

• Cause: a retry-exhausted failure is reported to checkContextDrift() at the same turn number the next real response will carry, because a failed call doesn't advance the turn. That analysis took the turn's slot, and the next response then hit the interval check.
• Effect: after any failed call (a 429 is the usual case), the next response was scored by none of the 23 signals, while its CDD_TURN still said analyzed:"true".

The fix:

• With exclude_infrastructure_errors on (the default): the infrastructure error returns before any analysis, and before the event-feed watermark moves. It feeds no signal, so nothing is lost.
• With it off: the error still feeds repeated_failures, so the analysis runs as before, then hands the turn's slot back.
• analyzed label: it now comes from a per-handle analysis counter instead of comparing turn numbers.

3. Untrack examples/hivemind/src/hivemind-telemetry.jsonl (10 MB)

It's the live output file of the hivemind configs, so every run appended to a tracked file, and nothing reads it. It's added to .gitignore; local copies are untouched.

Test Plan

• New tests/governance_v4/test_coherence_reconcile.sh: 7/7, registered in run-all-tests.sh.
  • RC-01: all 12 analyzed rows reconcile (worst residual 0.0001, from 4-decimal rounding).
  • RC-02: the run exercised every adjustment kind.
  • RC-03: negative control; without the new field, 11 rows fail to reconcile.
  • RC-04: penalties_detail stays signal-only.
  • IA-01/IA-02: under both exclude_infrastructure_errors settings, a verbatim repeat sent after a failed call fires response_repetition.
  • IA-03: the same repeat with no failure before it fires too, so the fixture can fire S21 at all.
• Against the One implementation of string slice/substring/replace; check inline code passed through process.run #262 build:
  • RC-01 and RC-02 fail: 11 of 12 rows don't reconcile.
  • IA-01 and IA-02 fail: the repeat escapes CDD while labelled analyzed=true.
  • IA-03 still passes.
• Leak check: 874/874.
• Still running, to be posted here: the related CDD suites (semantic signals, validation signal, signal contract, challenge fail path) and the full suite. test_event_feed.sh has already passed.

🤖 Generated with Claude Code

https://claude.ai/code/session_01ELUfjXZvx8kzXo1UJjrAhC

---

Generated by Claude Code

[github-actions]

NAAb Governance Report

Metric	Count
Files checked	16
Passed	16
Failed	0

All governance checks passed!

_{Generated by NAAb Governance Engine v4.0}

[b-macker]

Local results at 9aadb73:

• Full suite: 447 total, 381 passed, 52 error-behavior, 12 needs-tree-walk, 2 missing-executor, 2 unexpected.
  • The 2 unexpected are test_require_governance_gov007.sh and test_platform_fixes.sh.
  • Both are caused by a stray /govern.json on this machine, and One implementation of string slice/substring/replace; check inline code passed through process.run #262's run had the same two.
  • The newly registered test_coherence_reconcile.sh passed.
• Differential v2: 57 passed, 0 failed.
• Leak check: 874 passed, 0 failed.
• Related CDD suites, run directly:
  • test_semantic_signals.sh: 9 passed, 0 failed, 11 skipped.
  • test_validation_signal.sh, test_challenge_fail_path.sh, test_event_feed.sh: all pass.
  • test_signal_contract.sh: fails C3/C4 identically on the One implementation of string slice/substring/replace; check inline code passed through process.run #262 build. It's the documented acceptance gate that run-all-tests.sh deliberately skips ("expected to fail until the CDD signal redesign lands"). C1/C2 pass on both builds.

CI on 9aadb73 is green, including build-windows.

---

Generated by Claude Code