feat(autotrain): dual-regime thrash (isolate/climb/timeout residual)

[Tyler-R-Kendrick]

Summary

• Add pure thrash_regime helpers and policy v6 thrash regime routing (isolate / climb residual / timeout→decode residual).
• Record thrash_regime on matrix, cycle handoff, thrash timing, and SDLC delivery artifacts.
• Skip climb no-op residuals via materialized knob signatures; confirm/promote still freeze thrash.

Test plan

• pytest tests/test_autoresearch/test_thrash_regime.py + thrash/regime continuous tests
• version-stamps check for harness.autoresearch.experiment_campaign v182

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
slm-training	Ready	Preview	Aug 6, 2026 1:00am

[Tyler-R-Kendrick]

Rubber-duck + adversarial closeout review (agent)

Rubber duck

Walked the diff for PR #1455 in isolation vs its base. Intent matches the PR summary; docs + version stamps land with the code; tests pin the new invariants.

Adversarial

• Decode legality / I2: false-singleton path fails closed (incomplete domain) rather than inventing legality — does not weaken constrained decode or singleton bypass correctness.
• LALR bound: only negative-direction prune (room < certified lower bound → UNSUPPORTED). Certification required before use; failure disables prune for the session.
• CI weights: validator fails closed on partial / super-wall files without irreducible remediation; packing still degrades gracefully if table absent.
• Thrash regimes: production residual slugs stay constrained; confirm/promote freeze thrash; climb no-op skip uses materialized knobs not sparse overlays.
• Size / gates: no capacity growth; no ship-gate weakening.
• Coverage honesty: remaining topology live-apply and live dsh5 preflight stay @pytest.mark.slow — stated deferral, not silent skip.

No blocking findings. Proceeding with CI + squash-merge bottom-up.