feat: wire FDR into the gauntlet/discovery — multiple-testing complement over the whole search

FINDING (third this session): fxhnt's existing gauntlet DSR already uses the CANONICAL Bailey-LdP SE
((kurt-1)/4), while the Rust ml-validation SE uses (excess_kurt/4) — slightly non-canonical. So we do NOT
swap the DSR (the gauntlet's is already correct). The genuine add from the ported suite is PBO + FDR.
Wired FDR in: Verdict gains a pvalue (1−DSR, non-breaking default); DiscoveryService now FDR-corrects
(Benjamini-Hochberg at alpha=1−dsr_min) over ALL candidate p-values, and a candidate survives only if it
passes its own gauntlet AND survives FDR across the search. DiscoveryReport reports n_passed_gauntlet vs
n_survivors (post-FDR). PBO kept available for fold-based CV (future wire). 48/48 tests.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

src/fxhnt/application/discovery.py

@@ -23,6 +23,7 @@ from fxhnt.domain.backtest import run_backtest
 from fxhnt.domain.gauntlet import effective_trials, evaluate, per_period_sharpe
 from fxhnt.domain.models import Market, ResearchRun, StrategySpec, run_id
 from fxhnt.domain.strategies import get_strategy
+from fxhnt.domain.validation import fdr_correct
 from fxhnt.ports.data import DataProvider
 from fxhnt.ports.repository import AnalyticalStore, OperationalRepository
 
@@ -55,7 +56,8 @@ class DiscoveryReport(BaseModel):
     n_candidates: int
     n_effective_trials: float   # correlation-corrected independent trial count (the deflation uses this)
     sr_variance: float
-    n_survivors: int
+    n_passed_gauntlet: int = 0  # passed the per-candidate gauntlet (before across-search FDR)
+    n_survivors: int = 0        # ALSO survived FDR over the whole search — the final set
     survivors: list[ResearchRun] = Field(default_factory=list)
 
 
@@ -100,7 +102,7 @@ class DiscoveryService:
         # phase 2 — judge each candidate against the effective-search bar
         g = self._s.gauntlet
         now = dt.datetime.now(dt.timezone.utc)
-        survivors: list[ResearchRun] = []
+        judged: list[tuple[ResearchRun, object, np.ndarray]] = []
         for market, spec, prices, bt, r, split in rows:
             strategy = get_strategy(spec.kind)
             verdict = evaluate(
@@ -110,12 +112,24 @@ class DiscoveryService:
             )
             run = ResearchRun(run_id=run_id(market, spec), market=market, spec=spec, stats=bt.stats,
                               verdict=verdict, n_trials=n_trials_eff, created_at=now)
+            judged.append((run, prices, r))
+
+        # phase 3 — FDR over the WHOLE search: Benjamini-Hochberg on every candidate's p-value at
+        # alpha = 1 − dsr_min. A candidate survives only if it passes its own gauntlet AND survives FDR —
+        # the multiple-testing complement to the per-candidate deflated Sharpe.
+        pvalues = [run.verdict.pvalue for run, _, _ in judged]
+        fdr = fdr_correct(pvalues, method="bh", alpha=1.0 - g.dsr_min) if len(pvalues) > 1 else None
+        n_passed = sum(run.verdict.passed for run, _, _ in judged)
+        survivors: list[ResearchRun] = []
+        for i, (run, prices, r) in enumerate(judged):
+            fdr_ok = fdr.rejected[i] if fdr is not None else True
+            is_survivor = run.verdict.passed and fdr_ok
             if persist:
                 self._op.save_run(run)  # the verdict (cheap, SQLite) is kept for every candidate — audit
-                if verdict.passed:
+                if is_survivor:
                     self._an.save_returns(run.run_id, prices.dates, r)  # only survivors keep their timeseries
-            if verdict.passed:
+            if is_survivor:
                 survivors.append(run)
 
-        return DiscoveryReport(n_candidates=len(candidates), n_effective_trials=n_eff,
-                               sr_variance=sr_variance, n_survivors=len(survivors), survivors=survivors)
+        return DiscoveryReport(n_candidates=len(candidates), n_effective_trials=n_eff, sr_variance=sr_variance,
+                               n_passed_gauntlet=n_passed, n_survivors=len(survivors), survivors=survivors)

src/fxhnt/cli.py

@@ -234,7 +234,7 @@ def discover(
     report = svc.search(GridGenerator(markets, grids))
     typer.echo(f"searched {report.n_candidates} candidates  (effective independent trials "
                f"{report.n_effective_trials:.1f}, sr_var={report.sr_variance:.5f})")
-    typer.echo(f"survivors (deflated over the effective search): {report.n_survivors}")
+    typer.echo(f"passed the gauntlet: {report.n_passed_gauntlet}  →  survivors after FDR over the search: {report.n_survivors}")
     for run in sorted(report.survivors, key=lambda r: -r.verdict.dsr):
         typer.echo(f"  [PASS] {run.market} {run.spec.key():28} | Sharpe {run.stats.sharpe:+.2f}  DSR {run.verdict.dsr:.3f}  OOS {run.verdict.oos_sharpe:+.2f}")
 

src/fxhnt/domain/gauntlet/core.py

@@ -128,4 +128,6 @@ def evaluate(
         reasons.append("no structural rationale — refusing a purely statistical fit")
     elif has_economic_rationale is None:
         reasons.append("⚠ economic rationale UNVERIFIED — justify before deploying")
-    return Verdict(passed=ok, dsr=dsr, is_sharpe=is_sr, oos_sharpe=oos_sr, n_trials=n_trials, reasons=reasons)
+    pvalue = max(0.0, min(1.0, 1.0 - dsr))   # for across-search FDR correction
+    return Verdict(passed=ok, dsr=dsr, is_sharpe=is_sr, oos_sharpe=oos_sr, n_trials=n_trials,
+                   reasons=reasons, pvalue=pvalue)

src/fxhnt/domain/models.py

@@ -95,6 +95,7 @@ class Verdict(BaseModel):
     oos_sharpe: float
     n_trials: int
     reasons: list[str] = Field(default_factory=list)
+    pvalue: float = 1.0   # one-sided p-value (1 − DSR) — lets a search FDR-correct across candidates
 
     def summary(self) -> str:
         flag = "PASS" if self.passed else "REJECT"

tests/integration/test_discovery.py

@@ -40,7 +40,9 @@ def test_search_counts_full_matrix_as_trials(tmp_path) -> None:
     # every run was deflated by the same effective trial count
     assert len({r.n_trials for r in runs}) == 1
     assert 0 <= report.n_survivors <= 6
+    assert report.n_survivors <= report.n_passed_gauntlet   # FDR over the search can only tighten
     assert all(r.verdict.passed for r in report.survivors)
+    assert all(0.0 <= r.verdict.pvalue <= 1.0 for r in svc._op.list_runs())  # noqa: SLF001
 
 
 def test_effective_trials_below_raw_for_correlated() -> None: