feat(eval): record provider-resolved model id separate from requested alias

[hsskey]

Why

Evaluation run records stored only the requested model alias (for example opus), so a report could not tell which concrete provider model actually produced the scored output. That made cross-model comparison and later rescoring unreliable. This change records the provider-returned resolved model ID separately from the requested alias, keeps every legacy row valid and readable, and never lets rows with different model provenance share a comparison partition.

What changed

• The real Claude adapter runs with --output-format stream-json --verbose and parses the final result line for usage, cost, and text.
• LlmResult and EvaluationRun gain additive provenance fields: requestedModel, resolvedModel, modelResolution, per-call providerEvidence, and harnessCommit. One-shot rows also embed an artifact (output plus locked inputs) so a later scorer change can rescore from stored output without a new model call.
• Model resolution follows the SPEC 9.1 J3 rules: a producing-message ID when every main-loop assistant message names one identical model, a sole modelUsage key when no producing message exists, and null otherwise. Contradictory producing evidence resolves to null rather than falling back to a usage key, so a known contradiction is never hidden. No alias substring or cost heuristic is used as identity; raw provider evidence is preserved verbatim.
• model keeps its documented meaning (resolvedModel ?? requestedModel), so legacy rows continue to parse and group as before.
• One-shot report sections partition by prompt hash, model, requested model, resolved model, harness commit and dirty state, and reference-lock commit. null never merges with a value. Legacy rows lack the added parts and keep their existing sections.
• docs/reference/spec.md sections 9.1, 9.2, and 9.6 describe the new fields, the resolution order, the stream-json invocation, and the one-shot section key.

The trajectory partition, both trajectory reports, and the verify gate are intentionally untouched; this change records the row-level provenance fields only.

Follow-up commits on this branch: the explanatory comments this change had added were removed to match the change-history-only comment policy (behavior unchanged); and the one-shot section-key separator, which had been written as a literal NUL byte in the source (making the file diff as binary), now uses the same Unicode U+0000 escape as groupKey for the identical runtime separator and a text-clean source.

How I checked

pnpm lint
pnpm typecheck
pnpm build
pnpm tsx scripts/reference-lock.ts --check
pnpm verify --gate scope
pnpm verify --gate scoring
pnpm verify --gate types
pnpm exec vitest run packages/eval packages/verify scripts apps/cli
pnpm test

Results: lint, typecheck, and build clean; reference-lock --check reports no changed reference output; scope, scoring, and types gates pass; pnpm test passes 1335 tests across 75 files (net +18 provenance tests over the 1317 baseline). New tests assert resolved-ID persistence, the alias/resolved-ID distinction, partition separation for differing IDs, and legacy-row readability.

Full pnpm verify (all gates, includes --selftest) was then run from the branch tip: 12 of 13 gates pass (types, patterns, tests, reference, determinism, properties, mutations, rules, benchmarks, scope, scoring, encoding). The selftest gate failed once on a transient host-load timing collateral inside its own benchmark probe (allocation 1.842x vs the 1.5x limit, self-labelled environment: probe is slower than baseline); the top-level benchmarks gate itself passed. An isolated re-run of pnpm verify --selftest then returned PASS with 27 of 27 samples as-expected and zero failures, confirming the failure was an environment-timing flake, not a code defect. The provenance change does not affect allocation performance.

No incidental verify or benchmark artifacts are included: the run's appended bench/results.jsonl and its verify/*.json proof were reverted so the branch is the three provenance commits only.

What I left alone

No gate threshold, forbidden pattern, dependency allowlist, or self-test expectation was weakened. No committed run record, report, or benchmark result was rewritten, and no vitest -u was run. No real evaluation calls were made; all stream-json behavior is exercised against fixtures only. The trajectory partition, reports, and verify gate are unchanged.