[Engine] Split material identity from cell state behind a property table

[Kakapio]

Stacked on #13 — review that first; this PR's diff is only the material/property-table change.

Summary

Item 4 from the engine review: the biggest extensibility lever. The nested Particle enum tree (Common/Special(Ore|Gem)/Liquid(Direction)/Solid) conflated what a particle is with how it currently behaves, and adding a particle meant touching five-plus files (enum variant, sprite match, depth methods, spawn chance, interaction rules). This PR collapses all of it into a single registry. Net −157 lines.

What changed

Material + one properties() row per material

A flat identity enum whose entire static behavior lives in one const fn table: sprite index, Phase (Static or Liquid { viscosity }), and an optional WorldGenRule (Common depth band, or Special spawn weight + vein flag). Adding a particle is now one variant + one row + an atlas sprite. The generator, renderer, and simulator all read the table; the ParticleType/WorldGenType traits, the duplicated inherent depth methods on Common/Special, and the dead Liquid world-gen impl are deleted.

Particle { material, state } — honest equality

Flow direction moves out of the enum variants into CellState. Equality and hashing are now exact (derived), killing the footgun where Water(Left) == Water(Right) for HashMaps but not for pattern matching. Code that wants identity semantics — rule lookups, change tracking, deferred-step re-validation — now compares .material explicitly, which is what it always meant.

Directional interaction rules, no more HashMap

interaction_rule(source, target) is a const fn match on material pairs, replacing the LazyLock<HashMap> whose commutative key spun up two DefaultHashers per lookup in the hot path. Symmetry is now spelled out ((Water, Lava) | (Lava, Water)), and the asymmetric acid-into-water no-op — previously an accident of commutative lookup against directional Preserve semantics — is explicit and documented, with a note pointing at source-transforming rules as the future fix.

Phase-driven dispatch

Chunk::simulate picks a simulator by material.is_liquid() (the one dispatch point), and FluidSimulator reads viscosity from the table via let Phase::Liquid { viscosity } = …. A future powder or gas phase plugs in here without touching scattered match arms.

Behavior parity — verified, not hoped

• Both insta snapshots pass unchanged (byte-identical files in git): same-seed terrain generation is bit-for-bit identical (RNG draw order preserved — Material::iter() keeps Gold before Ruby), and the settled-pool simulation endpoint is identical.
• All conservation, interaction, sleep/wake, and determinism regression tests from [Engine] Overhaul simulation correctness, world generation, and chunk sleeping #13 pass untouched.
• Game smoke-run clean.

Tests

27 passing. New in this PR:

• sprite_indices_are_unique_and_nonzero — the atlas-collision insurance the old scattered indices never had
• particle_equality_is_exact — documents the new equality semantics
• Directional rule tests (water+lava both ways, water→acid, no-rule pairs)
• common_bands_cover_all_depths — terrain generation can never panic on a depth gap

🤖 Generated with Claude Code