Cross-grammar discrimination experiment (exp-002) + rotated-sampler fix

[jamesdward]

What

Milestone 2, Phase 7 — stacked on #2; I'll rebase onto main once #2 merges so this shows only the Phase 7 diff.

• Sampler fix (SI-017 gap 2, resolved): the rotated fragment crop degenerated to zero-size windows near 90° (negative slice index). The warp is now framed directly onto the window, making zero-size output impossible at any rotation; regression tests at 85–95° and extreme aspect jitter. exp-001's thin 90° arm was caused by this.
• battery/cross.py: four arms — 001 fragments and 002 fragments swept over fraction × rotation against BOTH sheets; a white-balance gain sweep on 002; and mutual impostors including a nine-ink grid impostor (002's exact structure, inks hue-rotated 165°).
• experiments/exp-002-cross-grammar/: manifest-pinned, 1269-row CSV, 4 charts, full report.

Headline results (from the report)

• Cross-grammar: 0 of 885 genuine fragments ever reached candidate against the wrong sheet (max cross-aggregate 0.268 vs true-sheet 0.35–0.88). The report is explicit that this separation is between-family and structural; same-family distance remains the open hard test (SI-022).
• White-balance (the M2 core claim): absolute colour matching collapses to 0.016–0.125 under casts; the relationship path holds 0.588–0.806 (above candidate throughout) and recovers the applied gain to two significant figures, degrading predictably in the strong-warm clipping regime.
• Peak discounting confirmed at its sharpest: an impostor sharing all of 002's canonical structure (grid, rhythm, staircase, exact multiply) but not its inks scores 0.003 — the address is the inks, and only the inks.
• 002 fragment strength: identified from 5% of surface area (vs 001's whole-surface requirement) — audit 002 §5's "dramatically better" confirmed empirically.

Tests

18 new (sampler regressions + quick end-to-end). Full suite: 125 passed.