Summary

Fixes gbrain#1819 and unifies all chat-model pricing into one canonical source.

The bug (#1819): Opus 4.8 (models.aliases.opus) had no price on file, so cost guardrails silently failed — the gbrain dream budget meter let runs proceed unbounded (BUDGET_METER_NO_PRICING warn + skip), and gbrain skillopt --max-cost-usd fell back to a cheaper tier's rate and refused too late. Adds Opus 4.8 at $5/$25 so the caps enforce again.

The unification: prices were hand-copied across five budget-gating tables and had drifted — one eval gate priced Opus 4.7 at a stale $15/$75 (3x high), and Gemini 2.0 Flash disagreed between two tables. All chat-model prices now live in one canonical table (src/core/model-pricing.ts); the Anthropic bare-key view, the takes-quality fail-closed allowlist, the contradictions cost-tracker, the cross-modal eval panel, and the skillopt preflight all derive from it. Cross-table drift is now structurally impossible.

Commits:

• fix(pricing) — canonical CANONICAL_PRICING + canonicalLookup; derived views; consumers rewired; Opus 4.8 added; stale Opus 4.7 ($15/$75→$5/$25) and Gemini 2.0 Flash ($0.30/$1.20→$0.10/$0.40, verified) fixed; brainstorm/brain-score now price provider-prefixed ids.
• docs — KEY_FILES.md entry for model-pricing.ts, CLAUDE.md "one canonical chat-pricing table" invariant, llms bundle regenerated.

Test Coverage

COVERAGE: 30/35 (86%)  — gate PASS (target 80%)
Fully covered: canonicalLookup (bare/colon/slash/nested-openrouter-miss/null),
  the drift guard (every derived view == canonical), opus-4-8 reachable across
  consumers, takes-quality fail-closed allowlist, S1A provider-prefixed pricing.
5 gaps: warn-only estimate paths + pre-existing untested consumer sites
  (orchestrator/skillopt/cross-modal estimate fns) — none touch the fail-closed
  gate or the drift guard.

Tests: new test/model-pricing.test.ts (integrity + canonicalLookup matrix + drift guard + S1A + import-cycle guard); updated anthropic-pricing.test.ts (opus-4-8 math) and eval-takes-quality-pricing.test.ts (opus-4-7 regression + opus-4-8). Full suite: 13,216 pass / 0 real failures.

Pre-Landing Review

SHIP — no CRITICAL. Enum/value completeness verified (opus-4-8 traced through every consumer). Applied doc-only auto-fixes: honest drift-guard framing (the guard is a re-hardcode trip-wire — drift is structurally impossible since views are derived), corrected a stale skillopt header comment, fixed a test-path reference.

Design Review

No frontend files changed — design review skipped.

Eval Results

No prompt-related files changed — evals skipped.

Scope Drift

Scope Check: CLEAN — diff matches the #1819 intent + the unification approved in plan review (D1A/D2B/S1A/S2A/S3A).

Plan Completion

25 DONE, 2 CHANGED (S1A used canonicalLookup instead of a literal prefix-strip — equivalent; correction #7 message updated), 0 NOT DONE. All locked decisions delivered.

Adversarial review (always-on)

• Claude subagent: SHIP. No bugs. Confirmed no OpenRouter-markup mispricing and no lost keys.
• Codex: informational (never gates). Its "block" rested on training-era pricing (claimed Opus should be $15/$75 — contradicted by the repo's 2026-05-10 verification and skillopt: add Opus 4.8 to anthropic-pricing.ts (cost-cap silently no-ops) #1819), a pre-existing+intentional OpenRouter-miss (markup ≠ native, by design), and a display-layer staleness in gbrain providers recipe baselines (not consumed by any budget gate). None are regressions in this diff. Real follow-ups filed as TODOs.

TODOS

• Completed: eval-contradictions duplicate ANTHROPIC_PRICING consolidation (this PR).
• Filed (P2/P3 follow-ups): route budget-tracker through canonical for non-Anthropic; canonicalLookup case-insensitivity; getPricing id normalization; module-load-throw negative test; recipe display-layer pricing refresh.

Documentation

Docs synced to the canonical chat-pricing unification:

• docs/architecture/KEY_FILES.md — added src/core/model-pricing.ts (canonical table); rewrote the anthropic-pricing.ts entry (was "single source of truth" → now a derived view); added the takes-quality derived-allowlist entry. Current-state only.
• CLAUDE.md — new cross-cutting invariant: one canonical chat-pricing table; every other table derives from it, so drift is structurally impossible (pinned by test/model-pricing.test.ts).
• docs/eval/SEARCH_MODE_METHODOLOGY.md — fixed a stale price-snapshot pointer to the canonical table.
• llms-full.txt — regenerated (build-llms freshness test green).

Closes #1819

This PR delivers the fix from gbrain#1819 (add Opus 4.8 pricing) plus the pricing-source unification that prevents the underlying drift class.

Test plan

• Full unit suite: 13,216 pass, 0 real failures (the 1 reported fail was cli.test.ts torn-reading VERSION/package.json mid-bump; passes 18/0 in isolation post-bump)
• Targeted pricing tests: 88 pass; consumer suites (budget-tracker, takes-quality runner, brainstorm, skillopt): 343 pass
• bun run typecheck clean; bun build --compile clean
• Runtime sanity: estimateMaxCostUsd / canonicalLookup / getPricing all return correct $5/$25 for Opus 4.8

🤖 Generated with Claude Code