Cache D3D12 recording lookups without retaining game resources

docs/windows-game.md

@@ -53,6 +53,8 @@ command-list split/replay, submitted state publication and suffix completion.
 The recording interval measures NGX calls on the recording thread, not Present.
 `-ProfileKernels` enables synchronizing HIP-event kernel profiling and changes
 frame scheduling; use it separately from a production FPS measurement.
+`-ProfileCommandHooks` samples one in 64 D3D12 interceptions to distinguish
+tracking/lock waits from driver calls. It is disabled in ordinary runs.
 The runner drains stdout/stderr concurrently as raw bytes; the earlier
 line-oriented PowerShell collector imposed large per-frame delays. The optional
 ZLUDA kernel profiler is for diagnostics only. Measurements and reproduction

docs/windows-performance.md

@@ -138,5 +138,39 @@ The installer accepts them with `-LocalTextureKernels` only after validation,
 checking the exact DLL identity, object and manifest hashes. M retains its
 independently validated FP16-equivalent baseline.
 
+The following no-debugger K/4K run completes 9982 frames and 9982 finite GPU
+scans. Kernel event profiling is disabled; the validated private output tail
+executes natively. The user reports 43-47 FPS. Mean NGX recording interval is
+21.867 ms. No backend failure, CPU image copy or previous-frame output occurs.
+
+Random one-in-64 hook samples then identify contention in the shared routing
+table used by every intercepted D3D12 command. Per-thread weak lookup caches
+now bypass that table on repeated calls to a list. Table mutations invalidate
+cache generations; Reset/address reuse cannot route to a previous suffix, and
+thread caches do not retain game resources. The per-list and submission locks
+remain. A GPU regression resets and re-splits the exact same list with changed
+root constants twice per iteration, checking both prefix and suffix results.
+
+| Sampled command-hook access | Shared table | Per-thread weak cache |
+| --- | ---: | ---: |
+| Busiest recording thread, mean us/call | 0.203 | 0.066 |
+| Estimated access + capture across threads, ms/s | 977.528 | 233.606 |
+
+The latter estimate includes lock waits summed across concurrently running
+threads, not CPU execution time or serial frame latency. The two game runs
+have different thread populations (31 versus 29); there is no identical frame
+trace. Generated forwarding methods have separate driver/capture markers;
+special-case hooks contribute access samples but lack those body markers.
+
+With the cache, 4770 complete K/4K frames and finite GPU scans pass, with zero
+backend failures, CPU image copies or previous-frame output. Median NGX
+recording interval is 19.473 ms. The user reports 49-51 FPS and 53% GPU usage
+in the previously measured scene. These user readings are separate from the
+sampled hook estimates and do not establish isolated WMMA utilization.
+All sixteen CTest gates pass, including 64 same-object Reset/split checks in
+each physical command-backend test and the software indirect-root regression.
+Use `-ProfileCommandHooks -ProfileStages` to collect samples; the sampler is
+disabled for ordinary play.
+
 References: [HIP events](https://rocmdocs.amd.com/projects/HIP/en/develop/doxygen/html/group___event.html),
 [HIP occupancy API](https://rocm.docs.amd.com/projects/HIP/en/latest/doxygen/html/group___occupancy.html).

docs/windows-rdna4-port.md

@@ -54,6 +54,27 @@ public ZIPs exclude these objects. Install/restore with private overrides
 restores sentinel DLL/config byte-exactly, without duplicate slash/backslash
 manifest paths.
 
+Production K/4K with private output stores completes 9982 frames and 9982
+finite GPU scans; the user reports 43-47 FPS. Random per-hook CPU sampling
+then localizes contention to the shared command-list routing map. A per-thread
+weak lookup cache bypasses that mutex for repeated calls, invalidating on map
+mutations to preserve Reset/final Release/address reuse. Queue and per-list
+synchronization remain. The regression resets/re-splits the same object twice
+per iteration with different root constants and verifies both GPU outputs.
+All sixteen CTest gates pass. The cached game run completes 4770 K/4K frames
+and 4770 finite GPU scans, zero failures/CPU image copies/aged outputs. The user
+reports 49-51 FPS and 53% GPU usage. Aggregated sampled access/capture time
+falls from 977.528 to 233.606 ms/s across threads; it includes parallel lock
+waits and must not be treated as serial frame time. Median recording interval
+is 19.473 ms. `-ProfileCommandHooks` is optional and off during ordinary play.
+Detailed scope and timings are in [windows-performance.md](windows-performance.md).
+Post-cache full-pipeline checks also pass: three K/4K HDR frames and four M
+512x288 frames match their previously validated controls with max absolute and
+relative RGB error 0, all RGBA finite. Public archive validation checks all 63
+entries and their SHA256 against package.json; NVIDIA DLLs and private texture
+objects are absent. `package-windows-game.ps1 -ArchivePath <zip>` archives only
+the staged manifest, including on Windows PowerShell 5.1.
+
 Performance diagnostics (2026-10-01): the separate ZLUDA `b0161a4` runtime
 passes 32 PTX launches in each context and all four native K/M frames with
 bit-exact RGB against their controls. All thirteen exported patches (0002