Fleet-cohort driver analysis
Research-grade — validation in progress.
Pure-library Spearman rank-correlation driver finder via `src/lib/fleetCohortDrivers.ts` — across N similar stages in a fleet, ranks candidate input features by |ρ| against a chosen outcome metric, with operator-tunable cohort filters (basin / play / σh band / lateral band / vintage window) and a plain-English rationale chip per driver. NOT a causal model — rationale strings read 'correlated with', never 'causes'. Drives hypothesis generation that then routes to `/sensitivity-sweep` (OFAT) or `/design-copilot` (Pareto). Workspace route `/fleet-drivers` ships the cohort builder + tornado chart + ranked-driver table; sidebar entry under Predict (`isNew: true`). Distinct from `fleetLearning.ts` (per-(kind,key) shrinkage estimator for priors) and `mlSurrogate.ts` (quadratic response-surface OFAT). Drift-guard: `fleetCohortDrivers.test.ts` (10 tests). Memory: `mem://features/fleet-cohort-drivers.md`.
Partial-correlation / mutual-information variants to disentangle co-linear drivers; auto-cohort-suggestion (k-NN over `inputs`) so the operator doesn't hand-pick the cohort; one-click hand-off to `/design-copilot` that seeds factor bounds from the top-K ranked drivers.
No phase markers extracted from this row's prose yet.
- No published error band yet — use with engineering judgment.