feat: price a run from a versioned rate table

[SurefireStudios]

Refs #10. Covers the first two bullets; the per-run budget is deliberately left out, see below.

What was already here

More than the issue implies. packages/submission/src/pricing.ts already had a rate snapshot, an estimator that prices input and output separately, and ModelUsage.estimatedCostUsd already permitted a number. What was missing was a table keyed by model, and any way to tell later which rates produced a figure.

What this adds

• PRICING_TABLE, keyed by the model id a manifest already records, so a lookup needs nothing else, with PRICING_TABLE_VERSION covering the table as a whole. A rate change bumps the version rather than editing an entry in place, so a later change cannot silently restate what an earlier run cost.
• priceRun returns the cost and the version together, because they have to agree. A figure without the table that produced it cannot be checked later; a version without a figure says nothing.

A model the table does not price stays unpriced. My first attempt passed the looked-up rates with ?? undefined, which fell through to the estimator's default and would have priced an unknown model at Opus rates. That is exactly the failure the file's own doc comment warns about, so priceRun now returns null for both fields instead.

Two things I checked before touching anything

Pinned artifacts. The hashes in submission/reproduction-manifest.json are over canonical predictions, not run manifests, so populating a cost field does not invalidate the frozen evaluation. pnpm reproduce still reports PASSED (27 checks).

Historical manifests. Twelve tracked manifests carry modelUsage with exactly five keys, and ModelUsageSchema is .strict(). pricingTableVersion is therefore optional, so those still validate. They are records of what happened and are not rewritten; the field appears on runs made from here.

One move worth flagging

The rate table moved from @stateproof/submission into @stateproof/core. The runners that make the model calls needed it, and making agents depend on submission would have inverted the layering, since submission reports on agent runs. Core already owns the manifest schema whose fields this fills, so the fields and the only code that populates them now sit together. Submission re-exports the table, so its importers are unchanged and there is one table rather than two that can drift.

Tests

Ten in packages/core/test/pricing.test.ts, aimed at the property that matters rather than at arithmetic: a cost figure is never invented.

• input and output priced at their own rates, since pricing both at the input rate is the mistake the split exists to prevent
• an unpriced model, and a null model id
• a missing, NaN or negative token count
• a zero-token run priced as 0 rather than as unpriced
• every table entry keyed by its own model id, so a lookup cannot return the wrong rates

The scoring-integrity guard asserted estimatedCostUsd was null because the feature did not exist. It now asserts the two fields agree, which is the real invariant. The report line that read "Cost in USD is deliberately null: no pricing rule has been implemented" was true and is now false, so it is replaced.

Full suite 547 passing across 31 files, pnpm typecheck clean.

Not in this PR

The optional per-run budget that stops compilation before it is exceeded. It touches compilation control flow and needs a decision about what happens mid-suite when the budget runs out, which is a product call rather than a mechanical one. Happy to take it as a follow-up once you have a view.

Note

Low Risk

Overview
This pull request adds a model-keyed PRICING_TABLE and supporting types to the core package, providing versioned input and output rates for pricing calculations. The submission pricing flow now applies these rates, reports input and output costs separately, and includes the pricing snapshot information used for a run. Runners and scoring logic propagate pricing metadata, while core exports, manifest schema fields, and pricing/scoring tests are updated to cover the new behavior. Per-run budget handling remains unchanged and is not included.

^{Written by Gitzilla for commit 6c3fccb. This will update automatically on new runs. Configure in the Gitzilla dashboard.}

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.