feat(campaign): discriminative scenario selection (E2)

[drewstone]

Additive primitive for research claim E2: select holdout scenarios by DISCRIMINATION power instead of a random/balanced split.

Motivation — the OR benchmark is saturating

Run 7 measured ~75% tied holdout cells: most problems are solved optimally by the baseline AND every candidate, so those paired cells carry zero signal. A random/balanced holdout spends its budget on scenarios that cannot separate candidates, and a tie in a paired cell wastes it.

What this adds

src/campaign/scenario-selection.ts:

• scoreDiscrimination(signals, opts?) — ranks scenarios by how well they separate candidates. discrimination = variance of the per-candidate composite scores (higher spread = more signal). Flags tied when variance < 1e-9 && meanScore >= saturationCeiling (default 0.999): every candidate scored the same at the ceiling, i.e. a saturated tie with no signal. Deterministic sort: discrimination desc, then meanScore asc (more headroom first), then scenarioId asc.
• selectDiscriminative(signals, k, opts?) — returns the top-k scenario ids for a holdout, EXCLUDING saturated ties when enough non-tied scenarios exist. Falls back to the least-saturated ties only when non-tied < k. Throws if k < 1; returns all ids in discrimination order when signals.length <= k.

Deterministic by construction (no Date/Math.random). Exported from src/campaign/index.ts.

Tests

src/campaign/scenario-selection.test.ts (11 tests, all passing): high-variance ranks above low-variance; all-candidates-score-1.0 flagged tied; an all-equal-but-LOW scenario is NOT tied (headroom remains); custom saturationCeiling honored; ties excluded when enough non-tied exist; fallback to least-saturated ties when non-tied < k; determinism; k<1 throws; k>=n returns all in order.

tsc --noEmit clean, biome check clean.

Additive; tested. No changes to existing behavior.

[tangletools]

✅ Auto-approved drewstone PR — f3ee4519

This PR was opened by the trusted drewstone account.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: drewstone_author · 2026-07-07T17:25:23Z}