Survivor × context stacking sweep

Thesis (PARTIALLY CONFIRMED): The two surviving classical patterns in the repo — coil_breakout and bull_flag — should be re-examined per regime to surface where each lives. Bucket the existing event lists by VIX z-score, trailing-252d drawdown, and 12m-1m excess momentum versus SPY.

Code: research/survivor_context_sweep.py Report: research/survivor_context_sweep_report.json (n=2,130)

Headline finding: bull_flag × elevated VIX is the strongest conditional cell in either survivor.

The 20-40d edge in the "elevated but not panic" VIX bucket is 2-3x the unconditional edge, with hit-rate jumping from 0.54 to 0.65 on n=202. The panic bucket (vix_z ≥ 1.5) shows the same direction but n is too small to lean on.

coil_breakout × context is much flatter — edges are similar across all VIX buckets (+1-5 %/20d) and the deep-drawdown bucket (dd ≤ -15 %) is the worst cell at +1.3 %, opposite to how the contrarian cohort behaves. Coil seems to be a regime-agnostic trend-continuation signal, not a regime-conditioned one.

Verdict: Keep bull_flag × elevated-VIX (0.5 ≤ z < 1.5) as a high-conviction sub-cohort. The intuition: in elevated-vol regimes, bull flags signal real institutional accumulation against a risk-off backdrop, whereas in calm regimes they signal nothing in particular.