Close the sub-OA dead zone, flag stale CDD telemetry, silence structural S16

[b-macker]

Summary

Round 6 from the post-#93 live-run forensics (the T26 cliff to coherence 0.445 and the dead zone after it). Every device finding was deep-verified against code paths before implementation — one was corrected (the "healing suppression" is proportional healing + interval skipping, not a bug: the observed −0.085 = −0.10 S16 + 0.015 half-healing, matching behavioral_sequence.cpp exactly), and the engine gap was sharpened: challenges required engine-global level ≥ elevated; after a de-escalation, re-escalation needs elevated_sustained pressure samples (~4 turns at check interval 2) while the quarantine streak (max 5) accrues every send — the recovery ladder was structurally slower than the kill path.

Changes

Engine (agent_impl.cpp, governance.h, governance_config.cpp):

• circuit_breaker.step_up_on_inadmissible (default false — no behavior change anywhere): with OA enabled, a handle whose CDD coherence is below the OA threshold draws a step-up challenge on its next send regardless of the engine-global level. Cooldown and max_challenge_failures still apply — this adds a trigger, not a bypass of the rails. Ratchet: disabling mid-run is a loosening violation.
• CDD_TURN telemetry gains analyzed ("true"/"false"): false = interval-skipped turn whose coherence/signals_detail are stale state re-shown from the last analyzed check. This display artifact misled two independent forensic passes; per-turn forensics should filter analyzed:"true".

Living-script configs (base + extended):

• Developer disables structural instruction_conflict — the round-5 absorption cap converted it from permanently-absorbed (free) to permanently-paying (0.10/check), because our own CONTRACT lines ("do NOT raise…") supply the negation markers S16 keys on. This completes the audit the cap design requires; CLAUDE.md now states that obligation.
• inadmissible_history: commit → exclude (quarantined content stops poisoning subsequent turns; tightening direction).
• step_up_on_inadmissible: true.

Docs: round-6 validation table in regression-surface-analysis.md (confirmed/corrected findings incl. the healing arithmetic); CLAUDE.md synced.

Test Plan

• test_challenge_fail_path.sh 16 → 21 assertions: three failed validations stage coherence at exactly 0.55 with the level pinned NORMAL (all pressure weights zeroed) → challenge fires via the floor trigger → pass recovers to 0.75 → run completes; default-off control draws zero challenges; mid-run disable rejected by ratchet
• test_absorption_degenerate.sh 10 → 12 assertions: at check_interval_turns: 2, six sends split exactly analyzed true=3/false=3; every CDD_TURN carries the field
• bash tests/security/test_error_msg_leaks.sh — 874/874 clean; test_validation_signal.sh unregressed (18/18)
• Both govern.json files JSON-validated; developer-scoped edit verified against accidental sibling-agent changes
• Full run-all-tests.sh sweep — 441 tests, 0 unexpected failures; both extended stub tests ALL PASSED inside the sweep
• Live run on the originating device — expect: no S16 drain on the developer, and any sub-OA excursion draws an immediate challenge instead of a quarantine march

Related Issues

Follow-up to #87–#93.

🤖 Generated with Claude Code

https://claude.ai/code/session_018L3RcVMf1tLNsGeJtofWmS

[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}