Stage-2 trend gate (Weinstein / Minervini) as conditioning filter

Thesis (CONFIRMED for two cohorts): Stage 2 — 50DMA > 150DMA > 200DMA with all three sloping up over the prior 20 days — should raise the per-event edge of a survivor pattern by selecting confirmed uptrends, the way relative_drawdown and vix_regime_overlay are used elsewhere in the repo.

Code: research/stage2_gate.py Report: research/stage2_gate_report.json (n=137,531)

Edges reported are unconditional (no controls subtracted) since the script re-uses each source event's own fwd_rets field.

Two clean wins, one neutral, one no-op:

• coil_breakout gains a clean lift from Stage-2 (+4.7 % vs +1.9 %/20d, hit 0.58 vs 0.50). The non-Stage-2 portion is what drags the unconditional average down.
• vcp is rescued from null by Stage-2 (+1.25 % vs -0.18 %/20d). Restores the Minervini context — VCP geometry only matters when the trend backdrop already says "uptrend".
• bull_flag is unchanged by Stage-2 — the pole-and-channel geometry already implies a local uptrend, so adding an MA-stack requirement is redundant.
• pocket_pivot still null inside Stage-2 — confirms the rejection above.

Verdict: Adopt Stage-2 as a stacking gate on coil_breakout and vcp. Skip for bull_flag (no lift) and pocket_pivot (null in both buckets).