P1: Turn research memos into portable rerunnable experiment bundles

[OnePunchMonk]

Outcome

Make the output useful enough that researchers share a report and collaborators can inspect or extend the experiment.

Current evidence

Reviewed main 38091c8c on 2026-09-13. Source: src/agent/research_memo.py.

The memo generator already exists and preserves unavailable fields. README says existing-manifest replay remains deferred until episode results are persisted as their own artifact.

Source review only: no new training run, benchmark or external pilot was performed for this issue.

Scope and dependencies

Delivery slice of #28 P3; extend the existing research_memo/episode_report implementation.

Work

1. Persist episode outcomes and references to all attempted candidates, configs, memory snapshots, dataset/split hashes and software identities in a versioned bundle.
2. Support report regeneration from saved outcomes without rerunning agents; keep this distinct from fresh execution. Include a rerun command and declared data-access requirements.
3. Add one user-data ingestion recipe with schema validation, timestamp/order checks and explicit cost assumptions. Keep private/raw data out of shared bundles unless explicitly included.
4. Export a self-contained HTML/Markdown decision report: hypothesis, evidence, baseline/candidate diff, unsuccessful candidates, outcome coverage and limitations; share existing renderer semantics.

Acceptance

• A second clean checkout regenerates the report from the bundle without model/network calls.
• An optional fresh execution validates data identity before running; absent private data is actionable rather than silently replaced.
• Two user-owned corpora/data fixtures exercise ingestion; malformed timestamps or missing fields fail before research.
• Every displayed measured number resolves to stored evidence; changing a report's presentation does not consume the final holdout.