perf(core): cheaper JSON byte accounting

[chh-ay]

Scope

A regression check of 0.5.0 against 0.4.x, plus one measured core optimization. Part of #59. Stacked on #72 (base bench/evidence-protocol).

Regression check

Matched 10-round render capture on the current harness, 7f73dc7 (0.5.0 stack) against fac48bd (0.4.x). The committed render baseline could not be used: it pins older harness sources and an older runner, and #72 re-records it in the release capture.

• No regression above A/A noise. The one strict-gate flag, arrow-right.middle (+5.8%, 99% lower bound +0.004 ms), did not reproduce in a 10-round attribution run at ≥250 ms per sample (+1.5% against an A/A spread of −7.0% to +5.7%). Rolling back the row-number clipping fix (b54b29a) did not change it.
• Editing scenarios lean slower but stay inside noise one by one: edit-open +3.0% to +11%, edit-commit +9.1%. scroll-right −14.7%, altering.remove-5-rows −9.5%.

Change

perf(core): JSON byte accounting (transaction admission) counts plain ASCII runs together, sizes finite numbers with String() instead of JSON.stringify, and checks own toJSON with Object.hasOwn. Byte counts, limits, and invalid-input errors are unchanged.

Workload	Before	After	Change
Clear + undo of 500,000 cells	935.9 ms	868.9 ms	−7.2%
CSV import (untouched)	447.2 ms	448.6 ms	+0.3%
Packed setBlock admission (untouched)	977.9 ms	985.2 ms	+0.7%

A/A control: 9 rounds, noise band ±4.6%. ABBA: 9 rounds; every round improved by 5.3%–9.3%. Outputs are identical. A faster XLSX attribute scanner was rejected (−1.7%, inside noise).

Verification

• Build, docs:check, api:check, typecheck, lint pass; packages 869 pass / 6 skip / 0 fail; Rust 147 / 223 pass.
• WASM binaries are unchanged by this layer (TypeScript only).

[chh-ay]

Landed on develop in the 0.5.0 release squash: #112 (55cfa82).