fix(models): treat subscription SKUs as honestly $0

[avs-io]

Problem

The unpriced-models warning from #638 fires on subscription / flat-rate product SKUs whose cost genuinely is $0. The suggested model-alias fix maps those ids onto another model's per-token rate and invents spend.

The first cut classified the wrong ids: it treated metered codex-auto-review as $0, missed kimi-for-coding-highspeed (the SKU #968 was filed around), and certified a synthetic Warp id (warp) that no provider emits.

Closes #968.

Root cause

findUnpricedModels / isExpectedFreeModel only treated local-looking models, model-savings mappings, and exact zero-rate user overrides as "expected free". Subscription product SKUs look like missing LiteLLM rows. The built-in classifier then used screenshot strings instead of the ids providers actually write, and model-flat-rate --remove could not unsilence a wrong built-in.

Change

Third state alongside priced and unpriced: flat-rate / subscription-billed.

• Built-in classifier for real product SKUs: ClinePass auto-genius, Warp auto / warp-auto-*, kimi-for-coding-highspeed, grok-composer-*, plus Warp Auto / Grok Composer display aggregation keys.
• Dropped codex-auto-review (metered; Auto code reviews continue with Auto review disabled and consume main model usage openai/codex#32224; priced by fix(models): price Codex activity ids via the official underlying model #1056) and unsourced big-pickle.
• User hatch: codeburn model-flat-rate <model> (--list / --remove). --remove opts out of a built-in so a false positive can warn again without a release.
• Shared unpricedModelHint(): alias only if billed per token; otherwise model-flat-rate.
• Parse-worker snapshot carries user marks and opt-outs. Daily-cache hash includes them (one-time re-derive on upgrade).
• Existing warp-auto-* aliases stay priced; those calls remain in the coverage denominator.

User impact

Known subscription SKUs drop out of the unpriced warning. Remaining unpriced rows still get an alias path, plus an honest hatch when $0 is correct. A wrong built-in can be opted out with --remove.

Preservation

No new LiteLLM rates. No removal of warp-auto-* aliases. No parser / billing-provenance rewrite. No Keychain / Buzz / #868 / #916 / #998 work. No second Extra High.

Testing

• Focused: models (SKU class, Warp auto vs synthetic warp, kimi-for-coding-highspeed, auto-review unsilenced, built-in opt-out + snapshot + hash) and model-flat-rate CLI (user mark round-trip, built-in --remove / re-add)
• Sibling: warp-auto-efficient stays isExpectedFreeModel=false because it still has a billable alias

[iamtoruk]

Reviewed in depth with real-data runs. The third-state concept and the unpricedModelHint() consolidation are genuinely good — but on the corpus the issue was filed around, the classifier gets 2 of 3 SKUs wrong and silences one real metered surface:

1. Blocker — drop codex-auto-review (leaf entry and the /^codex auto review$/i display regex): OpenAI's tracker says auto review consumes normal model usage (Auto code reviews continue with Auto review disabled and consume main model usage openai/codex#32224) — it's metered; fix(models): price Codex activity ids via the official underlying model #1056 prices it. Keeping the entry here hides $390+ of real spend and contradicts a sibling PR.
2. High — the case Unpriced warning fires on subscription-billed models where $0 is correct, and its suggested model-alias fix would fabricate spend #968 was filed around is not fixed: kimi-for-coding-highspeed (3.4M tok, a real subscription SKU on a real corpus) is absent from the classifier and still fires the warning. Add it.
3. High — warp claimed covered but isn't: the observed Warp id in real data is auto, not warp; the test at tests/models.test.ts:1029 uses a synthetic {model:'warp'} no provider emits, certifying a case that doesn't occur. Match what Warp actually writes.
4. High — no user escape from a wrong built-in: isBuiltInFlatRateModel is unconditional; model-flat-rate --remove only touches config entries. A false positive permanently suppresses a real unpriced warning until a release. Give --remove power over built-ins (or a --force-priced list) — the failure mode is silent under-reporting of money.
5. Medium: getDailyCacheConfigHash() losing its early return invalidates every user's stored hash → one-time full daily re-derivation for all installs. Fine, but undeclared — say it in the CHANGELOG.
6. Medium: big-pickle is classified as a "ClinePass codename" but appears under provider opencode in real data — an unsourced assertion of exactly the kind this PR exists to prevent. Cite or drop.
7. Smaller: display-string matching (/^warp auto\b/i) inside a pricing predicate is brittle (and fix(models): resolve and merge raw ids in models report #1053 changes those strings — land this before fix(models): resolve and merge raw ids in models report #1053); parser.ts:3251 serve-memo key omits the flat-rate hash (benign, deserves a comment); stray unbalanced-paren comment edit in shouldWarnAboutUnknownModel.

Suggested landing order across the trio: #1056 (with its version fix) → this one (rebased per above) → #1053 (re-keyed).

[ozymandiashh]

Code review is otherwise positive: the effective auto-merged code passed 239 targeted tests and typecheck, and the opt-out/flat-rate behavior matched the stated cases. I am not approving or merging this head because it currently conflicts with main in CHANGELOG.md. Please rebase and preserve both changelog entries; I can do the final live merge check on the resolved head.