Recogniser for grammar 002: white-balance-robust ink-set matching

[jamesdward]

What

Milestone 2, Phase 6 — grid-family measurers for iso-002, wired in via a measurer registry keyed by structure.type (band → cascade measurers, grid → new ones; the image is measured once per family, and 001's behaviour is byte-identical).

• Grid measurers (recogniser/measure_grid.py): grid module by edge-projection autocorrelation (audit 002 §2's own method), flat-ink extraction, parent/product separation via the committed multiply arithmetic, overprint-residual verification (audit §5: verification, not identification), stripe rhythm, coarse primitive-mix classifier (measured but uncommitted per SI-008).
• The milestone's core requirement — ink relationships must survive white-balance shift: ink_set scores as max(absolute, relationship). The relationship path estimates a per-channel diagonal gain from a consensus cluster of sheet/measured ratios, excluding channel observations lost to clipping (≥250 / ≤5), scores clipped channels one-sidedly (measurement loss, not identity mismatch), then re-separates overprint products in corrected space. The estimated gain and clipping report are published in every claim's working.
• Measured behaviour: clean 002 render identifies at 0.909; warm cast (1.15, 1.0, 0.85) that clips two bright inks still identifies at 0.781 with the applied gain recovered to within 0.01/channel; absolute-only matching collapses to 0.055 under the same cast.
• SPEC-ISSUES: SI-020 (relationship-mode colour scoring is undefined in the spec; documents the v0 design and constants), SI-021 (limits of colour-only overprint verification).

Tests

19 new tests including the clipping regression (exact previously-failing case pinned). Full suite: 107 passed. 001 recogniser tests unchanged except one iso-002 test that asserted the pre-Phase-6 absence of grid measurers, updated to assert the new graceful cross-family behaviour.