Meta-finding: chart pattern survey, take 2

This batch of seven new pattern + two context-conditioning experiments extends the meta-lesson from earlier in this file:

• Unconditional new chart geometries (VCP, pocket pivot, rectangle, ascending/descending triangle, Wyckoff spring, NR-cluster) all reject. Five of the seven are flat-to-negative even before adding selection, well-powered (n ≥ 1k each, pocket_pivot at 134k).
• Context-conditioning is where the wins live. Two real ones surfaced in this batch:
  • bull_flag × elevated VIX (0.5 ≤ z < 1.5): +7.9 %/20d, +11.8 %/40d (n=202).
  • coil_breakout × Stage-2 trend: +4.7 %/20d (n=408). Lifts VCP from null to +1.25 %/20d too.
• Failed-pattern fade confirms the asymmetry already implied by failed_breakdown: only coil_failed works as a short (−1.7 %/20d, n=184). The H&S / double-pattern mirrors don't.

The rule continues to hold: context is the signal, the pattern is the trigger. Geometry without a regime overlay gives at most the average; the regime is what splits "average" into "edge" and "anti-edge."