perf(worldgen): per-world wonder profile — undo the #698–#709 hot-path cost

[marceld23]

Pure refactor of the terrain-wonders hot path — bit-identical terrain, verified by an FNV hash sweep (surface heights + extra bands + tunnel spans) over every planet type × two seeds × continents on/off, captured before the refactor and matched after every step.

What was slow

SurfaceHeight — the hottest function in the codebase (256 columns/chunk, ~17k-sample per-world calibration, rivers, previews, placement) — re-derived per COLUMN what is constant per WORLD: feature gates with ToLowerInvariant() string allocations, Noise.Hash world rolls, repeated planet-key hashing. Worse, the tunnel carver rebuilt its whole worm polyline (dozens of xorshift rolls) for every column of every tunnel-bearing cell (half the map), and sea stacks/hoodoos/cenotes computed a full RawSurfaceHeight anchor before checking whether the column is even inside the feature.

The fix

• WonderProfile, cached per world like the calibration (static dict + lock-free instance fast path): planet seed, terrain grain, hybrid thresholds, continent profile, lowered style and one boolean per feature family. Per-column code reads booleans and doubles only.
• Per-cell tunnel worm cache + XZ bounding reject in the span sampler.
• Distance-first ordering for the sea-stack/hoodoo/cenote probes (behaviour-preserving — all gates are conjunctive).

Measured

• Harness sweep: 2174 → 1472 ms (−32 %)
• Local fast tier: 7m05 → 5m31 (−22 %); expect the PR CI test job to drop well below its 17-minute feat(worldgen): terrain wonders — scars, continents, overhangs & the tunnel carver #710 peak.

closes #712

🤖 Generated with Claude Code