perf: integer layout units end to end (scale-late F26.6, D-254)

[thejustinwalsh]

The complete integer layout-units migration (docs/planning/integer-layout-units.md, D-254): cluster advances, line fitting, justification, and positioning on F26.6 integers under one round-half-up contract, applied scale-late.

• Slice 1 — dense font registry (lane-neutral, shaper −9,433 raw bytes)
• Slice 2 — F26.6 advance stream, authoritative integer fit at the quantized width boundary, chunk-64 summaries with byte-exact break parity
• Slice 3 — retained adjacency glyph stream + change-scoped invalidation: column-resize −7.2%, measure-query −5.5%, font-size −12.2% (all 3/3 interleaved rounds at 22k glyphs)
• Slice 4 — euclidean integer justification sharing the fit's Q16 expressions; distributed totals exact
• Slice 5 — evidence fold: lane A/B table and corpus re-derivation statement in docs/packages/text.md, exact-f32-extent ABI pinning harness, checkpoint benchmark record, D-254 row

Fixtures held byte-identical except the deliberate quantized-boundary re-pins recorded with their commits. Open after the migration: the measure-query 0.6–0.9 ms stretch and the resize p95-under-4-ms tail, both stated in the evidence.