perf(core): trim per-keypress waste on the focus and state path

[chiefcll]

What

Three behavior-preserving reductions on the path that runs for every focus change (so, every navigation keypress):

• States.has() no longer allocates on prefixed lookups. It built a `$${state}` template string on every call to let unprefixed queries match (has('focus') matching '$focus'). The hot caller is updateFocusPath querying Config.focusStateKey, which is already $-prefixed, so that second lookup could only ever match a doubled prefix that nothing produces. A charCodeAt(0) check now short-circuits it; the allocation happens only when an unprefixed query misses the direct lookup.
• updateFocusPath drops its per-press Set. The Set existed only to diff the previous focus path against the new one, never escaped, and the paths are a handful of elements. A linear indexOf scan does the same diff with no allocation or hashing.
• _stateChanged skips the style-resolution branch when there is provably nothing to do. this._undoStyles is left as an empty array once a state style has been undone, and empty arrays are truthy, so the branch ran forever after that even when no style matched any active state. It now requires a non-empty undo list or a matching state style; every sub-case of the skipped branch assigns nothing.

Why

Keypress attribution work on the row-mount jank found the focus/state machinery is a small share of a press next to row construction inside Solid's batched key handler. These are the pieces of that machinery that were pure waste either way, aimed at allocation/GC pressure on TV-class devices.

Measured result, stated plainly

On a desktop browser these changes did not move the per-press benchmark numbers beyond run-to-run noise: an A/B/A sequence (6 benchmark runs per arm) showed environmental drift larger than any effect. They are strictly less work with identical observable behavior, so the recommendation is to gate this on a device benchmark run rather than take it as a claimed win. One semantic edge case to be aware of: a state literally named with a doubled prefix (e.g. '$$focus') would previously match a has('$focus') query and no longer does; nothing produces such a state.

Verification

• pnpm test: 178/178 pass, repeated runs
• pnpm tsc clean, pnpm lint clean
• Behavior cross-check during benchmarking: _stateChanged call counts per press were bit-identical between baseline and this branch

🤖 Generated with Claude Code