[Feature]: [P2] Engine hot-path scan removal: pickX per-spawn full scan and evaluateStar gate re-scans

From the full repository audit (2026-07-21, findings AUD-015 and AUD-016, recommendation N-5; report at _bmad-output/planning-artifacts/reviews/full-repository-audit-2026-07-21/). Companion to #338 (the render-frame congestion memoization), which lands first. Backlog row engine-hotpath-scans.

Problems

1. pickX scans the entire unit list per spawned person (AUD-015). src/engine/crowd/trips.ts:57-66 builds a tile list from a full tower.units walk per call; it runs twice per spawned trip (origin + destination), up to 8 spawns per outer step, plus once per dwell expiry: up to ~16 full O(25k) scans per sim step, a leading slice of the measured 8.6 ms tick.

2. evaluateStar performs up to ~20 full unit scans per hour (AUD-016). src/engine/sim/star.ts:13-47 re-derives population and each star gate (hasOperational/countOperational/recyclingDemandMet) with fresh full scans across up to 3 ladder iterations, every hour, even when the star cannot move: a named contributor to the 53 ms hour-boundary spike.

Fix directions

1. A revision-memoized floor -> structural-tile index for pickX (the tower already maintains structure/lobbyTiles incrementally in register/unregister). Constraint: preserve tiles[Math.abs(seed) % tiles.length] ordering so spawn positions stay byte-identical; index insertion order must match unit-array order.
2. One shared per-kind operational-count pass (plus a single population read) for the gate ladder, or early-exit when the target equals the current star.

Exclusions

The broader per-tick baseline scans (AUD-017: accumulateWaiting + duplicate totalPopulation + spawnFloors sweeps) and the hour-boundary demand-map sharing ride the spec-onhour-boundary-cost activation, sized by the field data from #538; they are not this issue.

Acceptance criteria / required tests

• Golden-master determinism unchanged (spawn positions byte-identical; a test pins pickX results against the un-indexed implementation on a populated fixture).
• Star evaluation results unchanged across a milestone playthrough (existing milestones integration stays green).
• Measured tick cost on the sixseven_15 save recorded before/after in the PR.

Engine hot path: /gds-code-review.

Start with src/engine/crowd/trips.ts:57-66 and src/engine/sim/star.ts:13-47, then inspect the tower's incremental structure/lobbyTiles maintenance. Run the pickX determinism test against a populated fixture and the existing milestones integration; done means byte-identical spawn positions, unchanged star results, and sixseven_15 tick measurements recorded before and after.