§4.14 World-building facts delivered via #176 (5aa481d); this closes
the loop on #119: re-verified all three tarball SHA256s by fresh
download (recast d48ca01…/box2d ffec747…/jolt 6e069ee…) and re-ran the
in-tree proofs (navigation P6-03 ALL PASS, physics_2d 2D-PHYSICS-OK,
terrain P6-01 ALL PASS, golden 5BC9BD960DA41391) on upstream 9090be1.
Fixes the stale §4.11 item 6 pointer that still claimed Recast/Box2D
APIs were unexercised. No slice code.
Refs #119
Co-authored-by: hermes-agent <hermes-agent@local>
* parity: 105 wire GiFrameGraph resolve output into the shade pass
Part 2 of P4-07 (part 1 merged as #191): per-frame MeshSdf-brick bake
+ probe upload, mesh.frag resolve-to-shade indirect term. Cornell floor
tint now comes from graph-produced data bound in the frame; no CPU-baked
volume in the path (unit-scale baker, gain in UBO).
Fix-forward vs main: P5-02 helper extraction had swapped two -Z cube
corners (uncovered left triangle per back-wall face); restored. Both
goldens match unmodified; differential best=0.112245, ECS_RENDER_OK,
GAME_GI_BAKER_OK. No ABI change.
Closes#105. Closes#197.
* parity: 105 fix CI round 1 — MSVC C4251 guard + MoltenVK-safe probe fetch
- gi_baker.hpp: standard _MSC_VER 4251 push/pop (std::vector members in
ENGINE_CORE_API class, same as the other renderer headers).
- mesh.frag: texture() -> textureLod(..., 0.0) for the probe fetch.
Single-mip volume (NEAREST, maxLod 0) so lod 0 is exact; explicit-LOD
fetches are derivative-free, which MoltenVK requires inside the
divergent gain branch (implicit texture() fails MSL transpile with
undeclared probeVolSmplr). Lavapipe output unchanged (goldens still
match, differential best=0.112245).
* parity: 105 fix CI round 2 — skinned layout + probe image init layout
- Skinned pipeline shared mesh.frag but kept joints at set 0 (STORAGE)
where the frag expects sampler3D+UBO: incompatible layouts (lavapipe
intermittent segfault/hang in engine_forward_mesh, MoltenVK refusal).
Joints move to set 1; skinned pipe layout is [shared probe set 0,
joint set 1] (probe resources created on demand); record binds both.
- mesh_skinned.vert gains the frag's location-2 v_world output.
- Probe image transitioned to SHADER_READ_ONLY at creation (was
UNDEFINED until first bake while every draw binds it as sampled).
Full local ctest green; forward_mesh stable 6/6.
* parity: 105 fix CI round 3 — branchless probe fetch for MoltenVK uniform flow
* parity: 105 diagnose mac black frame — expose gi_probe_live
* parity: 105 fix CI round 2b — probe upload outside the open pass
Transfer work (barrier+copy) recorded inside dynamic rendering breaks
tile-based drivers (MoltenVK black frame). Upload now runs as a
fence-waited one-time submit; create-time transition seeds SHADER_READ
so first upload barriers from there. Lavapipe output unchanged.
* parity: 105 notes round 2b — mac black-frame fix
* parity: 105 fix CI round 4 — zero-fill probe volume, branch the fetch
Branchless fetch read the uninitialized volume even with gain 0
(NaN texels poison pixels on drivers without zero-filled fresh
memory — WARP golden diffs, MoltenVK black ON frame). Fetch is
branched on gain again (textureLod stays: derivative-free, divergent
safe); creation zero-clears the volume via the one-time submit.
* parity: 105 review fixes — stale-gain reset, align-safe copy, split helpers
CodeRabbit round: ON->OFF now pushes a gain-0 box (stale indirect term
gone; test asserts off-after-on bit-identity). WorldTransform rows copy
via byte loop (alignas(64) vs flecs columns under -mavx2). Probe create /
upload split into small helpers (make_probe_image/host/sampler_set,
seed_probe_image, write_probe_box/copy_probe_image, gather_brick_rows);
uploads skip when the box bytes are unchanged (steady-state identical
rewrite also avoids touching in-flight host-visible memory).
* parity: 105 fix MSVC C4702 — drop duplicated return
* parity: 105 cherry-pick P5-03 physics flake fix (zero-init RigidBody rows)
Windows CI run on this branch hit the known physics_bodies flake
(free-fall y ~= -14.2 from garbage is_sensor -> Jolt sensor): cherry-pick
a3e6e71 from parity/p5-03-animator-state-machine-c. Test-only change.
* parity: 105 Sonar new-code pass — arrays, auto, split helpers
S5945: C arrays -> std::array (+ .data() at Vulkan/C call sites) in
GiBake params/light, GiBrickRow, probe box shadow, Vk descriptor sets,
bake AABB, test camera basis. S5827/S1905: auto + drop redundant
void* casts. S6022: std::byte loop. S107: TraceCtx bundles trace args.
S1188/S3574: AllocProbeHostBuffer + SampleSdfTrilerp + gather_probe
extracted; test add_cube/render_frame -> shared helpers. S6004
init-statements, S5817 const (gather_brick_rows/bake), S6181 bit_cast,
S6164 pi_v, S6188 span bake params. C++20-legal only (no std::print).
* parity: 105 fix GCC -Wconversion — size_t array indices
* parity: 105 notes rounds 5-6 — review + Sonar log
* parity: P4-07 GI resolve shade CodeRabbit round (frame scratch, payload compare, bake cap)
* parity: P4-07 GI resolve shade round 2 (shadow reset, baker scratch, cap order, docstrings, drop physics hunk)
* parity: drop stray clangd index, ignore .cache/
---------
Co-authored-by: hermes-agent <hermes-agent@local>
Co-authored-by: Hermes <hermes@local>
- kTaaEdgeVarianceMax 0.5 -> 1 LSB (identical-input golden must
converge exactly; old jitter derivation replaced)
- _spv.inc regen comments name the real outputs
(fxaa_frag_spv.inc / taa_resolve_frag_spv.inc)
- P4-10.md: MoltenVK divergence entry + current acceptance run
(worst=0.0000), replacing the stale worst=0.3020 record
Native ABI 4.25 tail (minor 4.24->4.25, append-only): entity_remove
after the P3-04 field tail (remove @1280, table 1288B).
GenericSystems::remove (no-op when absent) + ApiEntityRemove
(drops the run_systems lane plan like ApiEntitySet). Codegen
EngineEntityRemoveFn added to required entries. Pins updated in
test_abi_layout.cpp + AbiLayoutTests.cs.
Managed: FieldValue codec (all 11 field types + ZYX Euler-quat),
SetFieldCommand (entity_get/set; LocalTransform via
editor_get_local/set_local), Add/RemoveComponentCommand (LocalTransform
structural), EditorSession World/HasComponent/FieldsOf (+ ComponentFields
readback over the ABI 4.24 out-param FieldDesc tail).
Panel: InspectorPanel — per-component Expanders, TextBox lanes,
CheckBox, color swatch + RGBA lanes, entity ComboBox, asset TextBox +
drop. One undo step per LostFocus/Enter. Add Component menu + Remove
entry. Headless acceptance in InspectorPanelGolden (2.5+Enter sets
position.x, Ctrl+Z restores, golden editor_inspector_lm);
SetFieldCommand unit tests; inspector ctest engine_inspector_edits.
Forward-ported from parity/p3-05-inspector-panel (31c673b, Oct-31 P3
stack) onto the P3-04 merge (41c8e21): native side rebased from ABI 4.5
(@424/432B) to ABI 4.25 (@1280/1288B); MeshRenderer mesh_id reads as u32
(the 8B asset-ref lane would overlap base_color @4 and describe
rejects it); EditorSession keeps the P3-08+ DropAsset/DeleteEntity and
public World; shell keeps the P3-12 dock layout.
Co-authored-by: hermes-agent <hermes-agent@local>
