Read docs/PROMPT_capability_ratio_test.md

# Prompt — capability test: can the engine discover a planted gene ratio? (isolated)

Build a **self-contained, known-answer capability test** (like the H1 fixture)
that plants a synthetic target defined by a *ratio of two real genes*, rigs it so
neither gene is useful alone, and checks whether the blind engine rediscovers the
ratio. This is an engine unit test — NOT a biological claim, NOT a new UI
objective, NOT a new dataset.

**Isolation (non-negotiable):** live as a single new file `scripts/capability_ratio_test.py`
(scripts/ is allowed to be biology-aware). It READS the existing processed matrix
and reuses the existing sealed map via `airgap.reveal` for the final check. It
writes NOTHING to disk, adds no dataset, no API route, no UI, does not modify the
sealed map or any parquet. Deleting the file leaves zero trace. Do not touch
`engine_v2/`, `dsl/`, `api/`, or `web/`.

## Why a small panel (important)
Two individually-uninformative genes give the search NO univariate gradient, so
finding them among ~20,000 columns is an impossible needle-hunt — that tests
search-at-scale, not composition. To test composition fairly, restrict the search
to a small controlled panel: the two planted genes + ~50 random decoy genes.

## Steps
1. **Load real expression.** `cohort = Load("processed")` (colorectal). Keep
   samples with complete expression. Work on the NAMED matrix here (scripts/ may
   see gene symbols).
2. **Pick a correlated gene pair A, B.** From highly-expressed, adequately-varying
   genes, find a pair with high positive correlation (e.g. Pearson r ≥ 0.6) — a
   ratio only hides the single-gene signal when A and B move together. (Search a
   few hundred random candidate pairs; take the first that satisfies the r
   threshold AND the step-4 checks. Make the pair configurable via CLI for
   reproducibility.)
3. **Define the planted target.** `zA, zB = zscore(A), zscore(B)`;
   `signal = zA - zB` (the log-ratio direction); `y = (signal > median(signal)).astype(int)`.
   y depends only on the A-vs-B balance, not on either level.
4. **Validity checks (assert before running — this is what makes the test
   meaningful):**
   - single-gene AUROC(A) and AUROC(B) each within ~0.5 ± 0.07 (individually
     ~useless). If not, pick another pair.
   - the ratio itself, `max(AUROC(signal), 1-AUROC(signal))`, ≥ ~0.9 (jointly
     strong).
   - Print all three so the setup is auditable.
5. **Build the search panel.** Columns = {A, B} + ~50 random decoy genes from the
   matrix. **Control checks (print):** best single-gene AUROC over the whole panel
   ≈ 0.5 (nothing wins alone), and AUROC of the plain mean of all panel genes ≈
   0.5 (averaging doesn't work). These controls mean any high score can ONLY come
   from composition.
6. **Anonymise + run blind.** `M = anonymise(panel_expression)` (reuses the sealed
   map; opaque IDs). Run `engine_v2.run_v2_pipeline(M, y, objective=<binary AUROC>,
   ...)`:
   - objective: `from engine_v2.fitness import V2Objective; V2Objective(target="msi",
     binary=True)` — used purely as a binary-AUROC scorer; the "msi" name is just
     the internal carrier, the values are the planted label.
   - `prefilter_n=None` (panel is already small; must not univariate-filter out the
     ~0.5 genes), `scalar_share_override=0.0` (force per-patient-score / Vector
     programs so the winner is a readable Combine, not an Associate scalar),
     `coherence_weight=0.0`, diversity on (`tournament_k=2, p_mutate=0.85,
     immigrant_fraction=0.10`), a generous budget (e.g. `population_size=200,
     n_generations=40`), and a small `n_permutations` (e.g. 100) — or skip the null.
   - Run a few seeds (e.g. 3) and report each.
7. **Grade against the known answer.** For each seed's winner
   (`result["winning"]`): `reveal(winning["gene_ids"])` → symbols. PASS if the
   winner's genes are exactly {A, B} (or contain both) AND `winning["holdout_score"]`
   ≫ the panel's best single-gene AUROC (say ≥ 0.80 vs the ~0.5 controls) AND the
   `program_repr` contains a `Combine(` with `protected_div` (or `sub`/`mul` — any
   two-gene interaction that recovers the planted relationship; note which).
   Print PASS/FAIL, the winner `program_repr`, the revealed symbols, the holdout
   score, and the control numbers, per seed. Print an overall verdict (recovered
   the interaction in k / n seeds).

## Interpretation to print at the end (verbatim-ish)
- PASS (recovered in most seeds): "The engine can discover a genuine two-gene
  interaction blind — so when a real target (HPV) yields only averages, that's
  because the biology doesn't need a ratio, not because the engine can't build
  one."
- FAIL: "The engine did not recover the planted interaction even on a small panel
  — a real limitation to fix (raise Combine rate / diversity / budget) before
  claiming the DSL composes."

## CONSTRAINTS
- Airgap intact: the engine sees only the anonymised opaque panel; symbols are
  revealed once at the end via `airgap.reveal` (bounded — the winner's genes).
  Nothing in `engine_v2` changes.
- Isolated: one new file under `scripts/`, read-only w.r.t. data + sealed map,
  no API/UI/dataset/registry change. Existing tests + runs untouched.
- The point is capability, not forcing the answer: do NOT tell the engine which
  genes matter or that the target is a ratio. Restricting to a panel and to
  Vector programs is a fair framing, not a hint (it still must find WHICH two
  genes and HOW to combine them).

## Checkpoint
- `python -m scripts.capability_ratio_test` prints: the planted pair's single-gene
  AUROCs (~0.5), the ratio AUROC (~0.9+), the panel controls (best single ≈ 0.5,
  panel-mean ≈ 0.5), then per-seed winner `program_repr` + revealed genes +
  holdout score + PASS/FAIL, and an overall verdict.
- No change to engine/API/UI/data; deleting the file leaves no trace.