feat: dataset adapters + stage runners + scrutiny + prompt registry (ready to run on real data)

[sebasmos]

What

Implements the remaining pipeline logic so the repo is ready to run once real data is staged. Everything here is fixture/mock tested offline (no API keys, no real data).

• Dataset adapters (real formats, tested against synthetic fixtures matching the raw layout): MedQA (JSONL), MIMIC-CXR (metadata + chexpert-label + report join, one Case per image), CheXpert (with the Support Devices flag), NIH ChestX-ray14, and an EHR resource-context loader for stage 5. Each maps into the shared Case schema.
• prompts.py: versioned PromptTemplate + PromptRegistry with the core committee/orchestrator/referee prompts; the run manifest pins the (name, version) used.
• experiments.py: the stage runners that wire existing functions, run_pilot (stage 0), run_solo_baselines + run_holes_test (stage 1), run_cascade (stage 2), all injectable and offline-testable.
• scrutiny.py: the stage-5 benchmark-scrutiny panel (planted defects, flaw-detection rate, coverage-gap estimate, pluralistic vs homogeneous comparison).
• interventions.py: the real-time referee driver (uses the pre-emptive hook) + CheXpert natural-cue validation (rank correlation of injected vs natural-device susceptibility).
• docs/PIPELINE.md: the end-to-end data-contract doc.
• Fixes the two stale test_datasets.py tests (adapters now fail loudly on missing data instead of raising NotImplementedError).

What still needs real data (not code)

Only the data downloads and a live Gemini key. The dataset adapters parse the documented formats and are fixture-tested, so an owner points raw_root at the credentialed MIMIC/CheXpert or the open NIH/MedQA and runs. Those verification steps are tracked in the open dataset issues (#45/#46/#47/#49).

Suite: 292 passed, 5 skipped, ruff clean.

[felipeocampoos]

Reviewed the code and tested a real merge against current main (this branch is 11 commits behind, so I didn't trust GitHub's "mergeable" flag alone — that only checks for textual conflicts, not behavioral consistency with the code that landed since).

What I checked:

• Read through datasets/medqa.py, scrutiny.py, interventions.py in full, skimmed the other adapters and experiments.py/prompts.py. Consistent with the rest of the codebase's style: careful edge-case guards (flaw_detection_rate returns nan on nothing-planted rather than a false-perfect score, natural_cue_validation explicitly rejects a nan correlation via corr == corr, the scrutiny moderator can only veto/intersect with what seats actually flagged, never invent a detection).
• ehr.py intentionally raises NotImplementedError for the stage-5 loader — documented as such, not a stray stub.
• No secrets, no hardcoded paths, no bare except:/TODOs left in the new modules.

Tested a real merge, not just the branch in isolation:

• git merge main --no-ff from this branch: merges cleanly (no textual conflicts).
• On the merged tree: pytest → 476 passed, 6 skipped, 1 failed — the 1 failure is test_validate.py::test_checksum_is_stable_and_matches_hashlib, confirmed pre-existing on main itself (Windows CRLF vs a literal-string hash fixture), unrelated to this PR.
• ruff check benchmaxxing tests on the merged tree — clean.
• On the branch as-is (pre-merge, before validate.py/#100's fixes existed), 293 passed / 4 skipped, matching the PR description's "292 passed" closely enough (env-dependent skip count).

One non-blocking note: the branch is quite stale (11 commits behind main) — worth an actual rebase/merge-commit before landing so the merge that GitHub performs is the same one I tested, not just "no conflicts."

Code quality and test coverage are solid — approving.