engine(macos-arm64): publish 0ec39c95 — normal-roughness pre-pass built-in, tonemap cinematic stack, render-thread CPU profile, R0 P5 crash/freeze/device-lost reporting + structured engine errors; GodotSharp regenerated; Windows handoff

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NCXU49xtHLSry4TQEzVN2m

.agents/handoffs/forge-windows-gpu-diet-r0-p5.md

@@ -0,0 +1,23 @@
+# Forge (Windows) — rebuild engine main at 0ec39c95 (GPU diet + R0 P5)
+
+Mac `engine/macos-arm64` is engine `0ec39c95` (2026-10-17). Windows is still `1e7be28b1`. The game pins `4.7.2`
+without a hash until Windows carries the same commit; then pin `4.7.2@0ec39c95` in ariki-game `project.godot`.
+
+## Build
+`tinqs pull` the engine, then the `windows-x64` manifest flags (MSVC, `platform=windows target=editor
+module_mono_enabled=yes`), regenerate GodotSharp from that binary, publish binary + console binary + GodotSharp here
+with sha256s, note in `manifest.json`.
+
+## New and NOT yet compiled on Windows (no toolchain on the Mac)
+- `platform/windows/crash_report_windows.{h,cpp}`, `crash_handler_windows_seh.cpp` / `_signal.cpp` (MiniDumpWriteDump,
+  pre-started writer thread, unhandled-exception filter), `os_windows_stack_sampling.cpp` (freeze watchdog stack walk),
+  dbghelp/psapi linked in every target (`detect.py`, `SCsub`).
+- `drivers/d3d12/rendering_device_driver_d3d12.{h,cpp}`: GetDeviceRemovedReason + DRED on device removal.
+- Verification checklist: engine `.agents/wiki/crash-freeze-and-error-reporting.md` (8 steps: main-thread crash,
+  worker-thread crash, stack overflow, native freeze, C# freeze, D3D12 TDR via `dxcap -forcetdr --gpu-breadcrumbs`,
+  `--disable-crash-handler` writes nothing, release template build links dbghelp).
+
+## Bench after the build (the pacer box)
+ariki-game `.agents/plans/gpu-frame-diet-ui-opaque-prepass-2026-10-17.md` and `render-thread-diet-2026-10-17.md`:
+run the Jungle/Main stands with `--gpu-profile`; the RENDER CPU PROFILE block now prints `_draw wall` and
+`end_frame (graph replay + submit)` — the split of the 12.6 ms render-thread floor that Phase 0 asked for.

README.md

@@ -31,10 +31,10 @@ into GodotSharp, so GodotSharp had to be regenerated — and must ship here alon
 
 | Platform | Engine | Terrain3D | GodotSharp shipped? |
 |---|---|---|---|
-| `engine/macos-arm64/` | **4.7.2** (`1e7be28b`) — GPU-driven MultiMesh culling Stage A+B+B2+C (HZB occlusion, default off) + static cache single-sample / hard-filter options + true Metal GPU profile + parallel shadow fill, on top of the shadow diet, `get_video_adapter_memory()`, device-lost diagnostics | built-in module | ✅ yes (`GodotSharp/`) |
+| `engine/macos-arm64/` | **4.7.2** (`0ec39c95`) — normal-roughness pre-pass built-in, tonemap cinematic stack, render-thread CPU profile, crash/freeze/device-lost reporting + structured engine errors (R0 P5), on top of everything `1e7be28b` had | built-in module | ✅ yes (`GodotSharp/`) |
 | `engine/windows-x64/`  | **4.7.2** (`b1bc0aef`) — upstream bugfixes + static shadow cache (no memory query yet) | built-in module | ✅ yes (`GodotSharp/`) |
 
-Mac is engine `1e7be28b` as of 2026-10-05 night (`--version` prints `4.7.2.stable.mono.custom_build.1e7be28b1`; same code as `b0daf759`, relinked to match the Windows pair so `required_engine_build` is pinned to `4.7.2@1e7be28b`): GPU culling Stage C (previous-frame depth-pyramid occlusion, default off), zero per-frame cull uploads, static-cache single-sample + hard-filter options (default = unchanged), true per-pass Metal GPU profile, parallel shadow instance fill, on top of Stage B2 CPU band skip for GPU-culled MultiMeshes (the x4.7 draw explosion of `gpucull veg on` is gone; the path still stays off by default) on top of everything 51ed21d5 had (shadow diet incl. tile budget / live caster size cull / inverted-contract switch, video adapter memory query, GPU device-lost diagnostics) plus GPU-driven MultiMesh culling Stage A (compute frustum cull + indirect draws, `multimesh_set_gpu_cull`) and Stage B (per-instance GPU LOD bands, `multimesh_set_gpu_lods`), real Metal GPU timestamps (`renderGpuMs`), staggered far cascades, fit-to-visible-depth (SDSM-lite, default off), the Metal argument-buffer copy-queue stall fix and the Vulkan breadcrumb barrier fix (#185). Windows is `b1bc0aef` (2026-09-28, 4.7.2 + static shadow cache) and gets all of the above when Forge rebuilds from main - do not pin `required_engine_build` before that. Windows is `b1bc0aef` (2026-09-28, 4.7.2 + static shadow cache) and gets the memory query when Forge rebuilds from main. The engine source repo stays private. Check `manifest.json` for sha256. The 4.7 Windows cutover handoff is resolved: [`.agents/handoffs/forge-windows-4.7.md`](.agents/handoffs/forge-windows-4.7.md).
+Mac is engine `0ec39c95` as of 2026-10-17 (`--version` prints `4.7.2.stable.mono.custom_build.0ec39c954`): `IN_NORMAL_ROUGHNESS_PASS` (heavy materials skip their texture work in the SSAO/SSIL pre-pass), the cinematic stack (CA, vignette, grain) inside the tonemap pass, render-thread CPU in `--gpu-profile`, and R0 P5 crash dumps / freeze watchdog / D3D12 device-removed reason / structured engine errors. Windows is still `1e7be28b1` until Forge rebuilds from engine main ([`.agents/handoffs/forge-windows-gpu-diet-r0-p5.md`](.agents/handoffs/forge-windows-gpu-diet-r0-p5.md)); the game pins `4.7.2` without a hash until both platforms carry `0ec39c95`. Earlier: Mac is engine `1e7be28b` as of 2026-10-05 night (`--version` prints `4.7.2.stable.mono.custom_build.1e7be28b1`; same code as `b0daf759`, relinked to match the Windows pair so `required_engine_build` is pinned to `4.7.2@1e7be28b`): GPU culling Stage C (previous-frame depth-pyramid occlusion, default off), zero per-frame cull uploads, static-cache single-sample + hard-filter options (default = unchanged), true per-pass Metal GPU profile, parallel shadow instance fill, on top of Stage B2 CPU band skip for GPU-culled MultiMeshes (the x4.7 draw explosion of `gpucull veg on` is gone; the path still stays off by default) on top of everything 51ed21d5 had (shadow diet incl. tile budget / live caster size cull / inverted-contract switch, video adapter memory query, GPU device-lost diagnostics) plus GPU-driven MultiMesh culling Stage A (compute frustum cull + indirect draws, `multimesh_set_gpu_cull`) and Stage B (per-instance GPU LOD bands, `multimesh_set_gpu_lods`), real Metal GPU timestamps (`renderGpuMs`), staggered far cascades, fit-to-visible-depth (SDSM-lite, default off), the Metal argument-buffer copy-queue stall fix and the Vulkan breadcrumb barrier fix (#185). Windows is `b1bc0aef` (2026-09-28, 4.7.2 + static shadow cache) and gets all of the above when Forge rebuilds from main - do not pin `required_engine_build` before that. Windows is `b1bc0aef` (2026-09-28, 4.7.2 + static shadow cache) and gets the memory query when Forge rebuilds from main. The engine source repo stays private. Check `manifest.json` for sha256. The 4.7 Windows cutover handoff is resolved: [`.agents/handoffs/forge-windows-4.7.md`](.agents/handoffs/forge-windows-4.7.md).
 
 Long-term: CI (Gitea Actions) builds on push and publishes here / to S3 —
 `docs/ai/plans/build-pipeline-architecture-2026-06-07.md`.

manifest.json

@@ -4,16 +4,16 @@
     "macos-arm64": {
       "binary": "engine/macos-arm64/tinqs.macos.editor.arm64.mono",
       "godotsharp": "engine/macos-arm64/GodotSharp",
-      "version": "4.7.2.stable.mono.custom_build.1e7be28b1",
+      "version": "4.7.2.stable.mono.custom_build.0ec39c954",
       "engine_repo": "tinqs/engine",
-      "engine_commit": "1e7be28b",
+      "engine_commit": "0ec39c95",
-      "sha256": "5facd3e614ead0f5bb93bcd49b468a7b8b4dd7f23e26a5aa81d7243645547c0a",
+      "sha256": "dcc7a6d183806d3a88d055c3d5c394f22d6b8e4f0750a897c99189b8f602707c",
       "target": "editor",
-      "built_on": "2026-10-05",
+      "built_on": "2026-10-17",
-      "built_by": "Ozan/Kraken via Claude (engine round 2 evening: Stage C + housekeeping, single-sample static cache, Metal profiler fix, parallel shadow fill)",
+      "built_by": "Ozan/Kraken via Claude (GPU frame diet + R0 P5 crash/freeze reporting)",
       "toolchain": "Xcode (Apple Silicon), SCons, VulkanSDK 1.4.350.0",
       "scons_flags": "platform=macos target=editor arch=arm64 module_mono_enabled=yes vulkan_sdk_path=~/VulkanSDK/1.4.350.0/macOS",
-      "note": "Mac. Engine 1e7be28b = b0daf759 code + one docs commit, relinked so macOS and Windows carry the SAME engine commit for the pin. Mac. Engine 1e7be28b = b0daf759 code + a docs commit, relinked so macOS and Windows carry the SAME engine commit for the required_engine_build pin (4.7.2@1e7be28b). Engine b0daf759 on top of 615e852f: GPU-driven MultiMesh culling Stage C (previous-frame HZB occlusion for the camera cull set, default off: multimesh_gpu_cull_set_occlusion / rendering/occlusion_culling/gpu_hzb) + housekeeping (one command buffer per MultiMesh, zero per-frame uploads, one dispatch per MultiMesh, RENDERING_INFO_GPU_CULL_DISPATCHES/INDIRECT_DRAWS_IN_FRAME = 11/12); static shadow cache single-sample mode (static_single_sample + directional_static_shadow_set_single_sample, default off) and static_soft_filter (default true = today's soft PCF; false = hard 2x2, measured -16 ms GPU at a village stand on M1 with the cache on); Metal GPU profile timestamps drain before sampling (per-pass table is true now); shadow passes fill instance data in parallel. GodotSharp regenerated (MultimeshGpuCullSetOcclusion, DirectionalStaticShadowSetSingleSample present). Smoke on Kraken: 0 wrapper errors, world + large village. windows-x64 is 1e7be28b1 too (Forge, d4cf0b7); pin set in ariki-game project.godot."
+      "note": "Mac. Engine 0ec39c95 on top of 1e7be28b: IN_NORMAL_ROUGHNESS_PASS spatial built-in (heavy materials skip texture work in the SSAO/SSIL normal-roughness pre-pass: terrain pre-pass 2.5 -> 0.6 ms on M1 Max); per-Environment cinematic stack in the tonemap pass (RenderingServer.environment_set_tinqs_cinematic: chromatic aberration, vignette, film grain); --gpu-profile prints render-thread CPU per label + _draw wall + end_frame; R0 P5: crash reports in user://crash/ in every build (Windows minidump + JSON sidecar, macOS JSON + backtrace), freeze watchdog (debug/settings/freeze_watchdog/timeout_sec, default 10), D3D12 device-removed reason + DRED, AgentLog structured error records (get_error_records / get_last_error_id / set_crash_context). GodotSharp regenerated (EnvironmentSetTinqsCinematic present). Smoke on Kraken: new API + AgentLog methods + shader built-in; the game ran the same commit (symbol build) in-world at 60-70 fps. windows-x64 is still 1e7be28b1 until Forge rebuilds from engine main (handoff forge-windows-gpu-diet-r0-p5.md); the game pins 4.7.2 (no hash) until both platforms carry 0ec39c95."
     },
     "windows-x64": {
       "binary": "engine/windows-x64/tinqs.windows.editor.x86_64.mono.exe",