fix(completion): use current artifact and proposal revisions

[drewstone]

Product defect

A GTM operator can correct an offer, overwrite a deliverable, or withdraw an approval. extractProducedState currently appends every historical version, so completion can still match text or approval from a superseded version. GTM's producer already emits stable output identities expecting the latest write to be authoritative.

Change

Use the existing normalization point, not a product-local deduplicator. Artifacts reduce by exact name/URI/id fallback; proposals reduce by proposal ID. The latest observation replaces content/status, including missing content or a rejection. Distinct paths stay distinct and first-seen identity order remains deterministic. Raw history remains with the caller.

No new API, runtime dependency, agent loop, or GTM policy. docs/produced-state.md explains the behavior change, ordering requirement, and evidence limits.

Verification

Executed the candidate TypeScript source through TypeScript transpilation and Node assertions: 13/13 focused checks passed. Reverting the map reduction to the baseline append behavior passes 5/13 and fails 8/13, including superseded content and approved-to-rejected proposals. Added the corresponding Vitest regression file.

The focused check is not a full package signoff. pnpm typecheck, build, verify:package, and the full Vitest suite were not run: this environment has Node/TypeScript but no installed repository dependency tree and cannot resolve the registry. No benchmark digest or release version was fabricated.

Scope

This establishes current observed state for an emission-ordered input, not a provider transaction or live sale. Tool names remain invocation evidence, not successful-effect receipts. Consumers must collect independently verified payment/delivery outcomes and preserve trace ordering. GTM should consume the normal published package release; do not copy this reduction into the application.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, add credits to your account and enable them for code reviews in your settings.

[tangletools]

✅ Auto-approved drewstone PR — 7835687a

This PR was opened by the trusted drewstone account.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

This approval is provisional. It rests on the audit running. If the audit cannot run — for example the CLI bridge rejects it — this approval is dismissed rather than left standing, so an unrun check never reads as a passing one.

_{tangletools · auto-approval · reason: drewstone_author · 2026-09-18T19:05:58Z}