feat(predict/evaluate): claude pair → claude-opus-5-5 (proc-v9) — HELD until after the long-conference release

[modelmirror]

⛔ Do not merge. Held until results/ot2026-longconf is minted (~2026-10-20). Draft; it is stacked on #1962, whose commit appears here until that PR merges.

Closes #1963

What changed and why

This moves the claude predictor and evaluator from claude-fable-5-1 to claude-opus-5-5.

• pricing.DEFAULT_MODELS["claude-code"] → claude-opus-5-5. No actor pins a model, so this is what the fleet runs.
• MODEL_RATES["claude-opus-5-5"] is $4 / $20 per MTok. ModelRate gains an optional cache_read_per_mtok, and Opus 5.5 carries its published $0.20. The uniform 0.1× multiplier would bill its cache reads at twice that. usage.json is never re-priced, so every Opus cell would carry a ~20% overstatement permanently. Every other model keeps the multiplier. A test pins the override.
• The claude-code-action pin moves to 9ca9355 in all five places. That commit is v1.0.231 plus a single version-only commit bundling Claude Code 2.1.280, the first CLI that accepts the model id. No tagged release bundles 2.1.280 yet. Rebase step: re-pin to the first tagged release that does.
• Re-bless → proc-v9: the claude predictor and judge digests are new, and the four codex/gemini digests carry forward with proc-v8's bless moment. Bless moments are step 2's forecast. FROZEN_SINCE is guessed late (2026-10-21).
• Draft freeze-record entry for the third supersession shape. ‹› placeholders mark what the rebase fills in: the landing date, the declaration it cites, and the action release.
• docs/budget.md: per event goes from $15–17 to about $10. The claude pair is priced by re-costing its 122 proc-v8 predictions and 27 gradings at Opus rates: ~$1.35 a prediction, $1.30–1.40 a grading. That assumes Opus spends the same tokens, and the doc says so.
• Tests that pin the default are updated. The config/predictors.yaml comment example is updated (not hashed). The prompts' MODEL_ID example is left alone on purpose, because it is hashed and editing it would move all six digests.

Artifact served: every claude prediction and evaluation from proc-v9 on, plus the ledger's cost figures.

Before this can merge (in order)

1. A declaration entry on main before 2026-09-28 (the long conference). The third shape de-counts; it is pre-registered only if a dated entry closes the proc-v8 counted window while its outcomes are unknown. Today 0 of the 122 cohort events have resolved. This is a separate, small PR; see claude pair → Claude Opus 5.5 (proc-v9): predict/evaluate model, pricing and budget #1963.
2. The long-conference release is minted (results/ot2026-longconf). After this lands, the live frozen scope no longer reproduces the release's figures.
3. Rebase and re-read:
   • fedcourts process-digest --all, and the step-0 grep on origin/main;
   • de-count counts: predictions on all three engines, and evaluations under all three proc-v8 judge digests (these will be large by then);
   • the re-forecast size;
   • re-pin the action to a tag, and re-check the vendor price page;
   • fold in any owed prompt amendments (claude pair → Claude Opus 5.5 (proc-v9): predict/evaluate model, pricing and budget #1963 item 5).
4. After the carrying promotion, run step 4: correct the bless moments and instant, then tag prereg/proc-v9.

Effect check (after the carrying promotion)

• gh workflow run integration-test.yml --repo ModelMirrorAI/fedcourtsai --ref main -f scenario=engine-actions-smoke -f engine=claude-code -f deploy-environment=prod is green, and the resolve step names claude-opus-5-5.
• The first claude predict/evaluate cells after it carry process_version.label: proc-v9 and usage.json "model": "claude-opus-5-5".
• uv run fedcourts usage-summary over those cells replaces the priced budget figure.

Review

stats-reviewer

• One blocker: the declaration was folded into this held entry, which would land after the outcomes. Resolved by splitting it out (step 1 above); the entry now cites it.
• Recommendations adopted:
  • "closed counted window" wording instead of "shakedown", which would contradict the release;
  • the release-first gate;
  • the evaluation count across all three judges;
  • the per-model cache-read rate;
  • budget figures labelled as priced;
  • the step-4 wording for both directions.
• Its alternatives are presented in claude pair → Claude Opus 5.5 (proc-v9): predict/evaluate model, pricing and budget #1963, not chosen here: evaluator-only via a model: pin on claude-baseline, or an unblessed Opus shakedown actor.

code-reviewer

• No blockers. Adopted: the pin comment and "release" wording, the predictors.yaml example, and the comment nits.
• It noted that effort is neither pinned nor hashed; this is left to the maintainer as claude pair → Claude Opus 5.5 (proc-v9): predict/evaluate model, pricing and budget #1963 item 6.

Gate

• Full test suite: 4,985 pass. One secret-scan timing test failed under parallel load and passes alone.
• ruff and mypy are clean, schemas have no drift, and zizmor is clean.

Not run: docs-reviewer and workflow-reviewer on this delta. #1962's workflow review covers the CLI pin. Run both at the rebase, since the text will change anyway.

This PR touches .github/workflows/ and config/, so it is maintainer-merged. It is also held; I will not merge it.

🤖 Generated with Claude Code