W-10: material diversity — presentation-only soil split + slope outcrops

[mrtlvmbt]

W-10 — material diversity: presentation-only soil split + slope outcrops

Part of the "hex diorama" program. Integration branch: render-r12-terragen-preview — branch
off it (W-9 just merged into it — you build ON TOP of the talus pass), PR back into it with
explicit --base render-r12-terragen-preview (GOTCHAS: gh defaults to main).
Lane: worldgen (v2/crates/world/** only).

Goal (user ask: "почвы различного состава")

The 512 all-ON map must show ≥7 visually distinct BARE materials with spatially coherent patches:
bedrock, sand, till, basalt, tuff, permafrost + the NEW soil variants below.

HARD CONSTRAINT — presentation-only (17-round consensus decision, do not violate)

MaterialId inside classify remains the SIM SUBSTRATE. The biome cascade (override_biome,
incl. the Fertile branch) and caps_from/material_mult read the UNCHANGED substrate. Touching
the substrate silently rewrites caps/O2/NO3 ecology (Fertile dies for renamed cells; Bedrock
(0,1) collapses caps on steep-wet cells). The split lives ONLY in the presentation byte that
WorldView::surface_material returns.

Mechanism

A pure POST-classify pass that rewrites ONLY the surface_material byte in WorldFields, from
(substrate, moisture, slope) — all three already computed inside classify_and_caps:

• (a) Soil -> {SoilDry (new u8 discriminant 9), Soil, SoilWet (10)} keyed on the ALREADY-COMPUTED
  per-cell moisture at thresholds picked from a MEASURED moisture histogram @512 (Phase-0 below).
  Thresholds land as named consts.
• (b) Exposed rock/outcrop: where slope (height drop to D8 receiver, already computed in classify)

  = OUTCROP_SLOPE_THRESHOLD and the presentation byte would be Soil*, presentation byte =
  Bedrock. Landform-primary bytes (Basalt/Tuff/Till/Sand/Water) KEEP priority — never overwritten.

• Gated by the SAME any-landform-ON gate (OFF path byte-identical).
• Determinism: integer-only (no_float_guard), fixed iteration order, no hashing containers.

Phase-0 (measure FIRST, local run authorized — pure worldgen, <1 min)

On the POST-W-9 field (the staged seam classify_and_caps_staged exists — use it), @512, seeds
1 and 2: moisture histogram (deciles) over Soil-substrate cells; slope histogram over Soil cells;
pick (a) thresholds so the three soil classes each form visible, coherent regions (report the
class shares — they inform, they don't gate) and (b) the outcrop threshold from the slope tail.
Numbers table goes in the PR body; thresholds are pinned FROM it.

Tests (in-suite)

• ON-path INVARIANCE: biome, caps, resource arrays byte-identical pre/post the W-10 pass
  (the pass may only change surface_material) — @dim=64 and @dim=256 landform-ON fixtures.
• OFF-path byte-identity (whole WorldFields).
• Presentation-byte sanity: every cell's byte is a valid discriminant; landform-primary cells
  unchanged by the pass (targeted asserts on till/basalt cells).
• Patch coherence smoke: on the 256 fixture, SoilDry/SoilWet each form at least one 4-connected
  patch of >= 8 cells (no salt-and-pepper) — thresholds from Phase-0 make this achievable; if
  not, re-pick thresholds, do NOT relax the test.
• Existing goldens: conserved + w2/w5 must stay byte-identical (they hash biome/caps/materials on
  the OFF path). If any landform-ON golden vector hashes surface_material and moves — two-pass
  re-pin via CI, report honestly.

Acceptance

• @512 all-ON, 2 seeds: >=7 distinct presentation materials PRESENT with coherent patches (patch
  count / mean size reported); NO share-floor KPI; NO tuning of vents/ELA/landform physics to
  manufacture coverage (anti-forcing — physical knobs are not render KPIs).
• Screenshot evidence: use the render harness (--standalone --dim 512 --seed 1 --screenshot,
  material color mode) BEFORE vs AFTER — attach to PR. VERIFY every PNG by opening it with the
  Read tool before claiming anything (GOTCHAS).
• ON-path invariance + OFF-path identity tests land and pass LOCALLY (run them via
  ./scripts/test-bar.sh targeted and QUOTE the passing output — a compile-check is not a test
  run; two CI round-trips were burned on this mistake in W-9).
• compile-check PASS (quoted, from v2/) -> push -> bash scripts/ci-report.sh green (or honest
  two-pass re-pin if a material-hashing vector moves).
• code-critic self-review on the diff vs THIS issue; verdict block as PR comment.

Out of scope

• WorldView trait changes; render palette work (that is R-14); substrate/caps/biome changes of
  any kind; new landform knobs.

Process

Branch w10-material-diversity off origin/render-r12-terragen-preview. STATUS-line contract;
status.md last. Do not poll CI in a loop — one ci-report.sh call per pass.