fix(dx12): reconcile resource states across submissions

[besmpl]

Summary

• track first-use and final resource states per command buffer, then reconcile them in submission order with queue-local preamble barriers
• retain preamble allocators, command lists, staging resources, and explicit destination COM references until a trustworthy fence or safe terminal teardown
• handle 2D-array and padded 3D buffer/texture copies with exact D3D12 footprints, including block-compressed rows
• make packed depth/stencil SRV, DSV, clear, and companion-plane transitions agree, with complete descriptor-allocation rollback

This replaces the narrower but unsound draft #259. The DX12 indirect-argument ABI correction remains isolated in #258.

Why

Recording-time currentState is not authoritative when command buffers are recorded independently and submitted later or in a different order. The queue now owns scheduled state, plans non-mutating reconciliation barriers immediately before each command buffer, and commits final state only after execution is handed to D3D12.

Verification

• CGO_ENABLED=0 go test ./...
• GOOS=windows GOARCH=amd64 CGO_ENABLED=0 go test -exec=true ./...
• focused Windows amd64 and 386 cross-compilation for hal/dx12 and hal/dx12/d3d12
• git diff --check
• two independent read-only reviews of state/copy/view correctness and queue/COM lifetime

Remaining validation

This was developed on macOS, so native Windows execution and D3D12 debug-layer validation have not run yet. The PR is draft until that hardware evidence is available.

[kolkov]

The architecture makes sense — recording-time currentState is theoretically unsound for independently recorded command buffers, and submission-ordered reconciliation is the proper DX12 model. The depth/stencil plane handling and copy decomposition show real DX12 expertise.

However, our current DX12 backend runs 11,410+ frames stable on Intel Iris Xe with maxFramesInFlight=2. The practical risk of replacing working state tracking with a 2586-line rewrite that hasn't been tested on Windows is significant. State transition bugs manifest as TDR crashes or silent corruption — not something we can catch with cross-compilation.

Before we can consider merging:

1. Windows testing required. Run with D3D12 debug layer enabled. At minimum: gg integration example for 100+ frames on DX12 without DEVICE_REMOVED. We can help — our dev machine has Intel Iris Xe.

2. Consider splitting — the view helpers fix (planeSlice SRV), copy decomposition, and state tracker could be independent PRs. Smaller pieces are easier to validate.

3. Preamble allocator pooling — buildPreamble() creates a new allocator per submission. For frames with PRESENT→RENDER_TARGET transitions (every frame), this adds allocation overhead. Consider reusing from completed frames.

The quality of the DX12 knowledge is clear — we just need native validation before moving forward.

@lkmavi — this is a large DX12 rewrite, but the copy decomposition patterns may be relevant to the Metal texture handling in #261. Your perspective welcome.

[besmpl]

@kolkov Thanks for offering access to the Intel Iris Xe machine. I have pushed the review changes at 97c722c, including completed-frame preamble allocator/list reuse, the copy-footprint fixes, and the depth/stencil view/state corrections.

I do not have access to native Windows hardware. Could you please run the remaining validation on Windows with the D3D12 debug layer enabled?

• run the gg/examples/gogpu_integration example against this PR head with GOGPU_GRAPHICS_API=dx12 GOGPU_RENDER_MODE=gpu GOGPU_DEBUG=1 go run .
• confirm at least 100 visible frames
• confirm there are no D3D12 debug-layer errors, DEVICE_REMOVED, TDRs, or visual corruption
• confirm clean shutdown
• if possible, report the Windows version, Intel driver version, adapter, tested commit SHA, and relevant logs

Cross-compilation for Windows amd64/386 and the full CGO_ENABLED=0 go test ./... suite pass locally, but I cannot substitute those checks for native D3D12 execution.

[kolkov]

Architecture is correct. Submission-ordered reconciliation is the proper DX12 model, and this PR implements it well. The depth/stencil plane handling, preamble pooling, and COM lifetime management all show deep DX12 expertise.

Preamble pooling addresses our previous review feedback — completed pairs returned to idle pool, capped at maxFramesInFlight, with fallback on reset failure.

Must-fix before merge

1. Five test failures in copy plan tests.

TestPlanBufferTextureCopiesSplitsArrayLayers, ...AlignsPlacedFootprintsByWholeRows, ...UsesBlockRowsForCompressedLayers, ...Keeps3DDepthInOneSubresource, ...Preserves3DRowsPerImageAsSliceStride — the want structs leave footprintWidth, rowPitch, copyWidth, copyHeight at zero but equalBufferTextureCopyPlans compares all fields. Please fill in the expected values.

2. Double planBufferTextureCopies computation.

copyBufferToTexture and copyTextureToBuffer call the planner twice — once for barriers, once for copies. Store the result from the first pass and reuse it.

Windows testing

We will run validation on Intel Iris Xe with D3D12 debug layer as requested — GOGPU_GRAPHICS_API=dx12 with gg integration, 100+ frames, checking for DEVICE_REMOVED and debug layer errors.

[kolkov]

Saw your fix in 2f5bd8e — test expectations filled in and double planBufferTextureCopies eliminated. The placement boundary improvement is a nice touch.

We'll run validation on Intel Iris Xe with D3D12 debug layer next and re-approve after that.

Issues addressed in 2f5bd8e. Windows tested on Intel Iris Xe — all pass, debug layer clean.