DKR-R 1.0.5 Beta 15 - Playtest 6 performance candidate

HISTORICAL BUILD RECORD
The original findings and limits below belong to the named earlier build.
For current instructions, use docs/README.md and CHANGELOG.md in the
source repository. These old results do not qualify later packages.

====================================================

IMPORTANT SCOPE
This is a bounded performance/diagnostics candidate, NOT a declaration that
all five approved phases or independent Camera & Scenery parity are complete.
The accepted Playtest 5 Windows and Linux packages remain the rollback point.
Every participant in an Experimental Rollback match must use this same build.

Changes in this candidate
-------------------------
* Routine five-second renderer reporting no longer drains the GPU/WSI queues
  or locks the render-target manager. Exact render-target inspection is an
  explicitly intrusive developer diagnostic, disabled by default.
* Owned draw-sidecar conversion reuses decoder-owned vectors and avoids the
  temporary per-frame hash-set nodes used for background matrix deduplication.
  Sidecars are still copied into the identity consumer before reuse. No GPU
  consumer is allowed to borrow live simulation RAM or temporary vectors.
* Graphics -> Camera and scenery now has an Experimental performance
  diagnostics checkbox. It is process-local, not a lobby/gameplay setting.
  It enables bounded cumulative timing/counter summaries in runtime.log.
* Timings separate metadata conversion, CPU upload waits, frame matching,
  render setup, and successful swapchain-present-call intervals. These are
  CPU-side observations, NOT GPU timestamps or physical monitor scanout.
* Interpolation history-reset counters distinguish first frame, scene/epoch,
  correction/generation, hidden frame and repeated/non-forward history.
  Requested/rendered display-frame and skipped-workload counters are included.
* Scene summaries include actual level/type, input owners, viewport count,
  finished-racer count and the three-player TT-camera selection, so the same
  race can be compared before/after a racer finishes.

Preserved safety and functionality
---------------------------------
The Playtest 5 boss/victory-scene fix, full-game startup, existing HUD layout,
widescreen, enabled texture packs, sound/music policy and online save isolation
remain in place. This pass does not alter stable online backends, prediction
depth, frame/input ownership, checkpoint hashes, scene confirmation or rollback
history admission. Mandatory CPU uploader waits, whole-workload/present
completion, error-path drains and resource/shutdown retirement remain intact.
No submodule, RecompiledFuncs or RecompiledPatches source was edited. RT64
changes are authored Patch Pipeline inputs applied to new build-owned copies.

Five-phase status / remaining work
---------------------------------
1. Baseline preservation and measurement infrastructure: implemented. Actual
   three/four-player user measurements are still required.
2. Proven avoidable reporting/allocation overhead: changes implemented.
   Magnitude of the improvement has not been measured on the user's session.
3. Independently local scenery construction: NOT implemented in this candidate.
   Preferences remain preserved but gated in Experimental Rollback. The CPU
   discards some multiplayer objects at scene construction and culls other
   geometry before display commands exist. Decoder-only edits cannot restore
   absent geometry; using local preferences in checkpointed RAM would break
   determinism. A new visual asset/construction representation is still needed.
4. Cutscene/correction pacing: diagnostic coverage implemented. No speculative
   relaxation of history, confirmation or resource gates has been shipped.
   Targeted pacing changes still require the new scene/correction timing data.
5. Windows/Linux compilation and packaging: this candidate is provided for
   user playtesting, not claimed as gameplay-qualified or regression-free.

Qualification
-------------
No game, launcher, private probe, CTest or packaged-runtime self-test was run
autonomously for this pass, in accordance with the user's testing instructions.
Compilation, Patch Pipeline application and non-executing package/source checks
are separate from gameplay qualification. Test sources may be compiled; a
successful compile is not reported as a passing executed test.

Suggested user check
--------------------
Enable Experimental performance diagnostics before starting a normal three-
player match. Use the same track, resolution, AA and 180 FPS target as before.
Play through a cutscene and let one racer finish while the others keep racing.
Close normally and share runtime.log from host and a client. Look for
[rollback][timing], [rollback][counter], [rollback][scene-budget],
[rollback][draw-budget] and [rollback][fps] summaries. They do not log friend
codes or network endpoints. Samples are cumulative over the process lifetime;
start a fresh process for a clean comparison. Disable profiling afterward.

Developer-only switches
-----------------------
DKR_ROLLBACK_PROFILE=1 enables the same bounded telemetry at startup.
DKR_ROLLBACK_INSPECT_RENDER_TARGET=1 re-enables intrusive target inspection
and its diagnostic drain; do NOT use it for frame-pacing measurements.
