[Bug]: [P2] Per-frame congestion() rebuild is a render-path full-tower scan; memoize spatialCongestionByFloor + functionalParkingSet

Rescoped P3 to P2 by the full repository audit (2026-07-21, finding AUD-014; report at _bmad-output/planning-artifacts/reviews/full-repository-audit-2026-07-21/). This issue's original "measure first" gate is satisfied, and the measurement moved it from a sim-side suspicion to a confirmed render-path defect.

What happens

TowerEngine.tick() calls this.sim.congestion() every frame (~60 Hz) at src/render/excalibur/TowerEngine.ts:363 (plus peakCongestionHotspot() at ~6 Hz from trafficHud.ts). Under the default v2 sim model that is spatialCongestionByFloor (src/engine/sim/congestion.ts:124-198), which is not memoized and per call:

1. rescans all tower.units with censusCount(u) per unit to build popByFloor (25k units on a real endgame save),
2. runs the un-memoized functionalParkingSet flood fill (src/engine/tower/routing.ts:135-162), whose own doc comment assumes its callers keep it off the frame path,
3. walks every transport span with stopsAt = skipFloors.includes(floor), O(span x skipList) per shaft.

Impact

Dominant per-frame CPU cost on large towers, and the primary driver of the audit's measured 11 -> 5 -> 2 FPS scaling across 40 -> 7.5k -> 25k units on the owner's real saves (measured with and without GPU confounds; the CPU share is confirmed). Every big imported 1994-era tower hits this.

Fix

Memoize spatialCongestionByFloor and functionalParkingSet on the key set the engine already maintains: (tower.revision, tower.mealOverlayRevision, outer-step token). Inputs are exactly layout (revision), censusCount/customersIn (every change bumps mealOverlayRevision), occupants (hourly), and rushFactor (clock-window step). In-repo precedent: Crowd.queueView's (step, revision) memo (src/engine/Crowd.ts:179-191). A beginStep-invalidated cache gives frame reads a snapshot at zero staleness cost; the once-per-hour call inside updateSatisfaction stays correct for free.

Acceptance criteria

• spatialCongestionByFloor and functionalParkingSet compute at most once per (step, revision) regardless of frame rate; a unit test pins the call count under repeated frame reads.
• Congestion values are byte-identical to the un-memoized implementation across a golden-master run (no behavior change).
• Before/after frame-cost measurement on the sixseven_15 save (25,217 units) recorded in the PR.
• The heatmap/overlay and traffic HUD read the same snapshot within a frame (no torn reads).

Required tests

Call-count memo pin (unit), golden-master equivalence (integration), and the existing perf gate.

Engine hot path work: /gds-code-review.

Start at src/render/excalibur/TowerEngine.ts:363, then read spatialCongestionByFloor in src/engine/sim/congestion.ts and functionalParkingSet in src/engine/tower/routing.ts. Review the (step, revision) memo precedent in src/engine/Crowd.ts:179-191 and the required unit, integration, and perf tests. Done means repeated frame reads compute once per step and revision, preserve golden-master values, and share one snapshot across the heatmap and traffic HUD.