Preserve scored terminal model failures

[lluisinthedesert]

What changed

• accept Prime context_length and max_turns terminal rows when they are error-free and carry final verifier reward/partial credit
• narrowly normalize deterministic ProviderError 400 context-window overflows to aggregate score 0 while preserving the raw native trace and error
• keep generic 400, invalid JSON, 429, 5xx, and network/provider failures rejected
• surface the native/display stop reason and model-failure outcome in imported rows and the private scorecard

Why

The previous agent_completed-only gate censored genuine model capability and context-fit failures. Corrected-v3 produced both scored Prime context-limit stops and deterministic Anthropic prompt-too-long failures. Those are benchmark outcomes, not missing transport attempts.

Evidence

• validated the classifier against the three actual Haiku 4.5 automation context-window traces
• npm run check: 1,077 tests, build, typecheck, 37 public skills, package smoke
• focused regressions cover scored context_length, normalized context-window ProviderError 400, and rejection of 429/503/invalid-JSON/network cases

No corrected-v3 data was imported or committed.

[lluisinthedesert]

Added commit 22996e9 for corrected-v3 provider availability evidence. Import configs may now declare provider_unavailable models with reviewed receipt and incomplete coverage. They render in a non-scoring lane with canonical_score=null and are excluded from aggregate rows, pass rates, costs, leaderboard, and Pareto. Import fails if an annotated model also appears in canonical scored traces. Nemotron 3 Super remains unscored; no corrected-v3 data was ingested. Full npm run check is green.

[lluisinthedesert]

Added the exact Poolside Laguna XS 2.1 context-window message shape in commit HEAD: ProviderError 400 containing Input length N exceeds the maximum allowed input length of M tokens. Both supplied real automation traces classify as context_window_exceeded with aggregate score 0 while raw errors remain untouched. Generic and nearby 400s remain rejected. Focused tests and full npm run check are green.

[lluisinthedesert]

Corrected-v3 integration hardening added in this commit: diagnostic audit configs can render visibly non-authoritative private scorecards but import-prime rejects aggregate publication; authoritative parity retains strict 100% incumbent calibration. Recognized transient 429/5xx call attempts are accepted only inside otherwise-scored terminal traces when a later usage-bearing call proves recovery; terminal/unknown transport failures remain rejected. Exact OpenAI OverlongPromptError context-limit rows remain score-zero model outcomes. All 10 Cedar scorecards built, all five parity imports calibrated 100%, full npm run check green.