# ADR 003: No ECS, no SoA, no hand-written SIMD at this scale

Status: accepted (Phases 6–8, grounded in docs/PERF_BASELINE.md).

## Context

Measured tick cost is ~15–25µs median (debug) over ~40 entities against a
16,600µs frame budget. The costs are latency-shaped (lookups, one small
allocation, branches over dozens of items), never compute-shaped.

## Decision

- No ECS/archetype storage, no SoA migration, no SIMD intrinsics.
- Optimizations stay at the lookup/allocation/branch level (index caches,
  squared compares, buffer discipline).
- Reopen thresholds (any one reopens the question, with a fresh profile):
  simultaneous movers > 16 (spatial index first, SoA with it), tick p50
  > 1,000µs sustained on target hardware, or entity counts > 1,000.

## Consequences

- `LaneBox`/`BayBox` stay AoS; `nearest_hazard_*` stays a linear scan;
  each carries a comment naming the threshold that would change it.
- Contributors proposing intrinsics or an ECS must attach a profile
  showing the current shape is insufficient.