fix(research): keep Campaign evidence processing in ACK

[proerror77]

Change

After ACK had already verified and settled a Campaign, a workstation wrapper downloaded the same bulk results again and blocked the next experiment on that copy. The ACK shell also downloaded Mission/ZIP artifacts that Rust settlement fetched a second time.

Pre-holdout settlement now requires the ACK round-readback cache, verifies it against an independently fetched Campaign result, and preserves the existing semantic, cost, holdout and ledger checks. It constructs a complete native metrics/MLP report before committing settlement. The aggregate stays in ACK/OSS and may grow with round count; a small transfer receipt declares its size and whether it fits the 4 MiB workstation ceiling. Large reports stay cloud-side instead of trapping an already-settled operation. The ACK controller publishes and reads back that report before advancing; an interrupted report resumes the same cache and settlement without another Job, bulk GET or fit. Workstation downloads are only for presenting the light report.

Canonical data-bearing commands and controller modes reject non-Linux hosts before work begins. Documentation and AGENTS correct the former workstation-freeze guidance. Native controller deadlines are capped by the remaining original task/grant window, signed root Job budget and eight hours. The required absolute task deadline is recorded create-once, cannot be extended on resume, is checked by init and is enforced by the ACK entrypoint even after late manifest application. The full flow review records removed transfers, checks that must remain, and cloud resource cleanup limits.

Issue relationship

Closes #1166

Validation

• Native end-to-end cache/report contract: remote Mission/ZIP objects unavailable, correct cache succeeds, wrong settlement hash and changed ZIP reject, report regeneration preserves exact bytes.
• Bounded regular-file/symlink cache test; expired/late/resumed controller deadlines, in-Pod timeout, report failure before settlement, and multi-MiB aggregate preservation; workstation host guard and existing metrics contracts.
• Controller recovery suite: report GET interruption preserves one Job and one settlement, adds no bulk GET, and resumes identical report publication. Existing learning, Study handoff, immutable input and tamper scenarios pass.
• All 25 CLI contract tests pass, including required original-deadline parsing. Kubeconform v0.7.0 strict validates all 44 resources in 19 files; the reference Job uses a schema-valid conservative integer placeholder.
• Scoped alpha-harness Clippy with warnings denied; shell syntax/ShellCheck, publication packaging/admission contract and diff checks.

Runtime impact and rollback

Merging triggers the configured artifact publication workflows. Use runner/controller images bound to the same verified source for future runs; no new real-market Campaign is submitted by this PR. Existing closed #1165 results stay at source 65e8196f6196b89100159d8dd684b439cc44d332; its one-time ACK summary correction ran in eight seconds and returned 435,560 bytes, without fitting models again.

The pre-holdout settle caller must supply its cache. The ACK OSS identity/transport map needs create-only PUT and GET for the exact sibling model-report.json; missing access blocks report publication, retaining the completed work for retry. Linux is an accidental-host guard, not ACK attestation. Job TTL does not remove arbitrary PVCs or node annotations; scoped cloud cleanup remains required. Revert this commit and publish a matching image pair if needed; preserve immutable results, receipts and active ledger ownership. No model mathematics, dependency, cost/trading gate or sealed-holdout authority changes.

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Advanced

Run ID: d7de4a1c-d6bf-4ac8-bb6a-92360c08435b

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-09-13T06:08:53.898042Z	6c06334	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 6c06334617

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".