Tiles holding a slope over 35 degrees are meshed at two vertices per cell
on a clamped Catmull-Rom surface that passes through every grid node, so
cliffs and ledges lose the facets of the heightfield grid while the
collision stays on the grid. The domain field exposes the smooth
elevation and a separable, per-phase-weighted block sampler with linear
edge extrapolation and a one-cell margin for continuous normals; the actor
samples tiles on the worker thread pool and builds meshes afterwards.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
A FogVolume covering the site carries a fog shader whose scattering albedo
follows the same spectral transmittance as the seabed and whose density
grows below the reference depth, so haze darkens and turns indigo with
depth around every object; the global volumetric density moves into the
volume and the far depth fog only covers what lies beyond 128 m. Sand
ripples are bent by a slow domain warp, blended with a larger ripple
scale and faded out between 40 and 60 m, which hides the 3.2 m tiling.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Strata gain a dip, a slow seeded warp that bends and notches the beds, and
a per-bed thickness and hardness jitter with ordered boundaries, so walls
never invert. BoulderField gains an angularity parameter: elongated
super-ellipse slabs with flat tops, sharp shoulders, random heading and
tilt, footprint area preserved. Both sites are re-baked with the new
settings; the coral heads stay rounded.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Rock is normal-mapped triplanar in world space at two scales, sand carries an
oriented ripple normal map on the xz plane, and a grain map drives detail
occlusion, albedo mottling and roughness. The three maps are seamless,
seeded, mipmapped textures built synchronously by SeabedDetailTextures, so
no asset is needed and builds are deterministic. The mesh has no tangents:
perturbed normals are computed in world space and written to NORMAL.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Strata now take a world rectangle, so the 4 m ledges cover the whole break
and talus south of z = 118 m without touching la Saignée or the landmarks.
Roughness gains octave and fractal-type parameters; a SlopeRoughness
modifier roughens rock faces only. Both sites are re-baked with hummocks
from 250 m down to 5 m. The secondary item, the first debris piece and one
safe-route waypoint move a few metres to keep their contracts.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Review findings: the mesa forced rock onto the canyon bed, two runners
declared the wrong TEST_COUNT, and the escarpment check could pass on the
slope alone. Canyon gains floor_material, the showcase is re-baked with sand
beds, the contract measures the step across the crest band, counts distinct
ledges, and probes the crevasse against the open plateau. Degenerate recipe
parameters now fail an assert.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The bake order is now primitive sum, modifiers (with the local gradient),
floor clamp, slope material, primitive materials, material patches. A
regression test checks the Banc du Vigilant still bakes byte for byte.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The site terrain integration now also configures zone 1 (3 s budget,
seabed under the hull, trench at the groove entrance, reference scan
from the spawn). The actor test uses an 81 x 65 grid with an
off-diagonal bump, checks tile vertices against their field cells and
samples the collision between vertices.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Building the 64 tiles of the Banc du Vigilant cost about 1.9 s through
per-vertex cell_world_position/normal_at_cell calls, leaving the 3 s
load budget with too little headroom (2.1-2.9 s measured, one 4.8 s
outlier). TerrainHeightField now returns positions and normals of a
cell block in one call, with normal_at_cell sharing the same formula;
the tile build drops to about 0.5 s.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
DiveSite.configure() builds the height field once (TerrainHeightField
validates the data) instead of running the definition validation and
then the field construction again; the composition root relies on the
error returned by configure(). The collision shape takes the field's
row-major elevations directly instead of copying 263k cells one by one.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The target hull now lies along the groove axis (yaw derived from the
groove constants), set back 22 m from the groove end so both hull ends
rest on the flat groove floor. The mooring, spawn and first route
waypoints move to z = -210 and the decoy yacht to (-30, -150) so no
wreck is farther than 320 m from the mooring and no two wrecks are
farther than 260 m apart (spec 4.1). The spawn faces south toward the
site. Four contract assertions pin the budget, the spawn heading and
the hull orientation.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Zone 1 becomes the default dive site; the foundation suite still selects the
calibration site explicitly. Documents the site suites, the site content
workflow and the authored heightfield terrain (ADR 0012), and aligns the zone 1
brief with the baked data (storage format, route lengths, 6 m ridge and
pinnacles).
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
The calibration site spawn sat 16 m from the south bounds edge, inside the
40 m warning margin. The vehicle_test recipe now extends the terrain and
playable bounds 42 m to the south so the spawn is inside the zone.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
chindit