Fizlet 0.4.1

Hardening from the engineering review (2026-10-29, §8: UI and tokens, collision and effects).

• pack smoke test: data subpaths (.css, .json) resolve instead of importing.
• ui: mount() keys its shared constructed stylesheet by the CSS text. It cached the first sheet for the whole page, so a second game mounted with different CSS got the first game's styles. Stylesheet ownership is documented (BOARD-UI.md 2.1). No visual change.
• ui: tokens.json is the authoritative token source. Published as fizlet/ui/tokens.json; the build fails when a board-ui.css default or applyPalette's variable map disagrees with it (scripts/tokens.mjs, tested). board-ui.css is byte-identical.
• collide: Grid separates entity capacity (cap), index capacity (new 5th argument nodes, default cap * 4) and query-result capacity (out), and every saturation is counted: dropped (index full; now counts every refused cell entry), rejected (id outside 0..cap-1 or a non-finite circle; such ids used to alias the dedupe stamps), truncated and missed (query results that did not fit; they used to vanish silently). The constructor and resize throw a RangeError on non-finite sizes or non-integer capacities.
• collide: add, near, insert and query pass numbers through a typed array, so the grid allocates nothing even when V8 does not inline it (the mixed collide + fx test went from ~24 young GCs per 200k frames to 0; its allowance drops from 100 to 1).
• fx: blast() stops at capacity instead of attempting every particle on a full pool (10 blasts of power 1e6 on a full pool: ~4.7 s before, ~0 ms now), and counts refused particles in fx.dropped. effects() validates cap, scorch (0 now turns scorch marks off) and the colour and word tables; spawn, blast and kick refuse non-finite input (a NaN life used to hold a pool slot forever).
• fx: no allocation in a live scene. blast() allocated on every call (Math.random returns a new heap number); the visual random is now a private generator seeded from Math.random (still never the simulation's seeded random) and blast writes the pool directly. Same recipe, same distributions.
• tests: the allocation test runs a populated fx scene (blasts every frame window, the pool full, drops counted) and asserts no young-generation GC; new regression tests for every item above, and a browser test for two games mounting different CSS.
• README: the overload policy (decorative particles may drop, counted; collision results never drop silently).