reproducibility-check

# Reproducibility check Stage 08 must write review and readiness artifacts under `workspace/reviews/`. The purpose is to state honestly what a third party could reproduce from this run's directory alone — not to assert that everything is fine. The output of this check is a record of what was **verified**, what was **not verified**, and what is **known not to reproduce**. All three categories are legitimate. Only a fourth is not: recording an unchecked item as verified. ## Audit the run directory as a stranger would Work from `runs/<run_id>/` and assume no access to the conversation, the operator's memory, or the environment the run happened in. | Question | Where the answer must live | | --- | --- | | What was the research goal? | `user_input.txt` | | What was actually run? | `workspace/code/` — scripts, not descriptions | | What data went in? | `workspace/data/`, with provenance for anything downloaded | | What came out? | `workspace/results/`, indexed by `experiment_manifest.json` | | How do the figures relate to the results? | a script under `workspace/code/` that regenerates them | | What decisions were made and why? | the decision ledger in the stage summaries | | Which stages actually completed? | `run_manifest.json` — see below | ## Read the manifest before claiming the run is complete `run_manifest.json` distinguishes stages that were **approved** from stages that were **skipped**. A skipped stage has `skipped: true` and a `skip_kind` of `human` or `auto`; an `...