refactor: dedup identical opus-tier rates in modelPricing EXACT_RATES (#4163)

[atomantic]

Summary

All five claude-opus-* rows in EXACT_RATES (server/lib/modelPricing.js) carried the same hand-copied [5.0, 25.0] literal, and the { test: /opus/i, rateModel: 'claude-opus-5' } entry in FAMILY_RULES had to be re-pointed at the newest id on every opus bump — migration 206 and every bump before it — purely to relabel the reported rateModel, since every opus row already resolved to the same numbers.

The tier now owns one OPUS_TIER_RATES pair plus a newest-first OPUS_MODEL_IDS list that generates its EXACT_RATES rows, and the family rule takes its pointer from the head of that list. An opus bump is now a one-line prepend — no second edit site.

Which fix shape, and why

The issue offered two shapes. I took shape 1 (shared constant, pointer derived rather than hand-written) instead of shape 2 (a family rule yielding rates inline), because rateModel is load-bearing downstream and shape 2 could not preserve it:

• client/src/pages/UsagePage.jsx prints Priced as ${m.rateModel} in the per-model tooltip.
• attributedModel() in server/services/usageReconciler.js compares two ids' resolved rateModel to decide whether a recorded launch model and a transcript model are the same family, and deliberately treats null as "not recognized, therefore not a match".

An inline-rates rule would have to report rateModel: null (breaking that comparison outright) or a synthetic label like claude-opus (which would stop matching the exact-table label, so opus and claude-opus-5 would no longer reconcile as one model). Keeping the pointer a real EXACT_RATES key avoids both while still retiring the per-bump edit — and it now cannot go stale or dangle, since it is derived from the list rather than typed.

No pricing or behavior change. A differential run of resolveModelRates and isFreeModelId over 954 provider/model pairs (every listed id, CLI shorthands, Bedrock-prefixed and suffixed variants, unlisted opus generations, local family:tag ids, and the fallback paths) returns byte-identical results before and after. No client mirror of this module exists — it is server-only.

Tests

server/lib/modelPricing.test.js gains an opus tier block pinning the three properties that made the hand-maintained pointer safe:

• every listed opus generation bills at the same tier rates (matched: 'exact');
• an opus id the table has never heard of (claude-opus-9, global.anthropic.claude-opus-7-…-v1:0) still prices at $5/$25 without a pointer bump — the invariant this refactor exists to guarantee;
• the reported family label is itself a non-null, exactly-resolvable table id at the same rates, and every opus id agrees on it — the property usageReconciler and the usage tooltip depend on.

The existing opus shorthand assertion now checks the rates rather than hardcoding claude-opus-5, so the suite needs no edit at the next bump either; the new label test covers what that assertion used to.

Test plan

• cd server && NODE_ENV=test npx vitest run lib/modelPricing.test.js services/usageReconciler.test.js — 89 passed.
• cd server && NODE_ENV=test npm test — 29127 passed. Four files reported failures under this run, all pre-existing load flakes from concurrent suites (a different, non-overlapping set failed on a second run); services/loras.test.js, services/imageGenQuota.test.js, routes/settings.secretsStrip.test.js, and routes/imageGen.watermark.test.js all pass in isolation (93 passed) and none touch pricing.
• Differential harness comparing pre- and post-change resolveModelRates/isFreeModelId across 954 pairs: 0 mismatches.

Closes #4163

[atomantic]

Review log — antigravity (agy), 2 rounds

Round 1 — 2 findings, both applied in e790d69:

1. Shared array reference. Object.fromEntries(OPUS_MODEL_IDS.map((id) => [id, OPUS_TIER_RATES])) handed all five opus keys a reference to one array, so a future edit to a single opus row would have silently rewritten the whole tier. Each row now clones the pair, matching the hand-written rows around it.
2. Weak test assertion. The family-label test asserted only expect.any(String). Tightened — but to toMatch(/^claude-opus-\d/) rather than the suggested hardcoded toBe('claude-opus-5'), since hardcoding the current head id would reintroduce exactly the per-bump edit this PR removes. The exact-resolution assertion immediately after is what actually pins the pointer to a real table key.

Round 2 — NO FINDINGS on the updated diff.

Re-verified after the fixes: lib/modelPricing.test.js, services/usageReconciler.test.js and lib/index.test.js pass (94 tests), and the differential harness still reports 0 mismatches across 954 provider/model pairs.