Fix model invocation attribution

[ppkulk-oss]

What changed

• Separate invocation evidence from performance evidence in the model scoreboard.
• Preserve requested and harness-reported model identities, with alias-aware matching.
• Register Fugu Ultra, current Grok models, and Claude Opus 5.
• Recognize Grok's selectable grok-4.6 model and reported grok-4.6-build usage alias.
• Attribute Claude from the primary assistant event's message.model, not the first modelUsage key (which can be an auxiliary Haiku call).
• Capture effective reasoning effort across Codex, Grok, OpenCode, and Claude invocation styles.
• Keep mismatched or unattributed rows unranked instead of silently crediting them.

Why

Grok runs without an explicit model were being stored with a blank identity and displayed as unattributed legacy rows. Fugu Ultra and Claude Opus 5 had no registry entries. Claude JSON can list an auxiliary Haiku model before the requested Opus model. Together, those gaps made the proven/probation table ambiguous because routing evidence and task-performance evidence were presented as one concept.

Impact

Future runs show whether an identity was harness-reported and matched, only configured/pinned, mismatched, legacy, or unattributed. Performance remains a separate proven/probation/unranked signal. Existing rows without defensible identity evidence are not rewritten.

The installed Claude lane pins claude-opus-5, explicit effort max, one attempt at the manifest level for the acceptance canary, and a four-turn CLI ceiling. It parses stream JSON assistant events, so auxiliary usage cannot be credited as the primary model.

Validation

• python3 -m unittest discover -s tests -p 'test_*.py' — 259 tests passed.
• Hardened Claude integration suite — 13/13 passed, including the fenced access-token-only OAuth-status check; no inference occurred in that suite.
• Live bounded Claude canary — PASS on attempt 1 in 2 turns.
  • requested model: claude-opus-5
  • reported assistant model: claude-opus-5
  • Ringer invocation: reported match · max
  • performance: probation (one task; not yet enough evidence for proven)
• Offline before/after Grok fixture:
  • before: unattributed / unranked
  • after: Grok 4.6 · high / reported match / proven
• Fugu launcher offline preflight passed.