feat(improvement): train and serve a checkpoint through improve with a receipt

[drewstone]

Status

Ready for review; all four required CI checks pass on final head 87694f15b5af2c669ffa844df63dc31485166127.

https://github.com/tangle-network/agent-runtime/actions/runs/35417723461

Passed: ci, agent-bench, official-optimizers, and packed-cohort. This includes the full test suite, lint, source/example typechecks, release-version checks, packed exports, edge execution, and generated-document freshness. Main 466a7a2237b0ed603159b4f02a96b90d628995be is integrated with no conflicts. The PR is non-draft.

The repository still requires one approving review. Review has been requested from the existing tangletools reviewer. The Codex review bot reports exhausted review quota; its unavailable review is not represented as a pass. No protections were bypassed.

Purpose and existing work

Train weights through improve(profile, { mode: 'training', ... }) or a bound harness's train(...), then return an exact checkpoint-backed candidate and portable training receipt. Training does not imply a held-out improvement or authorize promotion.

The reusable pieces already existed, but not this checkpoint-to-profile composition:

• Published Interface 2.10.0 owns receipt schemas, artifact-addressed model identity, and ancestry validation (agent-sdk#341/examples: supervisor + coordinator MCP, workers on sandbox OR cli-bridge (one knob) #342).
• Eval owns training export policy, evaluation, and promotion. Existing row payloads remain unchanged; Runtime inventories all exposed train/validation tasks and refuses cross-partition reuse.
• Runtime already owns runAbortable and publishExclusiveDurableFile; this PR now reuses them instead of maintaining parallel cancellation and publication code.

The original dependency blocker is resolved. Peer floor, catalog, frozen lockfile, and exact packed cohort agree on Interface 2.10.0. Public imports use the existing root entrypoint, not an unsupported ./improvement package subpath.

Correctness repairs

• Cancellation during input writes or checkpoint fsync cannot dispatch training or serving afterward.
• Checkpoint mutation during serving or candidate validation is rejected before profile publication.
• The receipt is durably published before the profile. Cancellation after committed publication does not retract the result. Storage failures attempt durable withdrawal and report cleanup failures explicitly.
• Failure results retain the failed stage and uncertainty about external training/serving resources.
• The command trainer pins its executable and declared inputs, supplies only an explicit public environment, bounds output, and cancels its POSIX process group.

Regression evidence

Re-ran the original PR implementation at ace732e0d7d118db899b2a819a513ea43f0d3912 with Interface 2.10.0 installed. Three new regressions fail there: training dispatch after cancellation, serving dispatch after cancellation, and accepting checkpoint mutation by the validator. Thus the dependency bump alone was insufficient.

The repaired product tree passes 95 focused tests across five files, including 29 training tests (19 original plus 10 new boundary regressions). Source typecheck and full lint also passed locally. Normal final-head CI above independently passes the complete repository checks.

Earlier clean preparation ran 4,226 tests successfully with six existing skips. Its temporary logging pipeline hid a Bench version-check failure behind tee; artifact inspection caught that. Bench's patch bump fixes it, and normal final-head CI now proves the release gate passes. The temporary preparation workflow was retired and is absent from this PR.

Release and boundaries

Prepare Runtime 0.242.0 and Bench 0.13.4, with changelogs, generated fixtures/API metadata, and packed-consumer import/type probes. Update this same PR by fast-forward; no force-push or mainline code loss.

No new optimizer, GPU scheduler, managed-training implementation, or Router deployment client. Command execution is trusted host code, not an OS sandbox. Managed adapters own remote cleanup and reconciliation. These checks prove execution/receipt contracts, not production model quality or a live GPU/Router deployment. No merge, publication, deployment, or paid model run performed.

Part of tangle-network/blueprint-agent#2473.

[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 — 74b00acb

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 · 2026-09-19T03:06:49Z}