DKR-R 1.0.5 Beta 15 - Playtest 8: local scenery/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.

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

This is a user-test candidate, NOT completion of all five graphics phases.
The accepted Playtest 5 ZIP and AppImage remain untouched as the rollback point.
Every participant in an Experimental Rollback session must use this build;
the owned CPU contract is checked at admission. Do not mix playtest versions.

Connect
-------
1. Every racer installs this same Playtest 8 package and selects their own
   unmodified US v1.0 Game Pak. Experimental v1.1/online mods are not enabled.
2. Host settings -> backend/synchronization -> Experimental rollback. Create
   a normal lobby; others join through Quick Join or a friend invitation.
3. Mark racers ready and start the ordinary countdown. The normal intro,
   title and menus should appear: no separate launcher or forced test track.
4. Use the normal overlay/controls. Host Prediction and Strict Input Sync
   remain available for the next lobby; don't change backend during a match.
   Single-player saves remain separate from the isolated online session save.

Late remote inputs use checkpoint restoration/replay in gameplay. Menus and
resource-changing transitions remain confirmed and retirement-gated; audio,
EEPROM and external effects remain confirmation-gated. Those safety checks are
not removed for performance. No ROM, save, profile or RAM fixture is supplied.

Implemented changes since the accepted baseline
----------------------------------------------
* Bounded RAM/observation pools reuse storage after ownership retirement. No
  rewindable live guest RAM is exposed to the renderer. Pool/queue limits and
  stale-work rejection remain enforced.
* Delta checkpoints update their comparison base from changed pages instead
  of recopying unchanged RAM. State hashes, full snapshots and rewind checks
  remain active. Successful checked-memory accesses no longer update failure-
  only diagnostics on their hot path; bounds/alignment/budget checks remain.
* Removed retail BHV_SCENERY sprite billboards are captured before deletion,
  retained as immutable visual resources, and drawn in native per-viewport
  passes. This targets the removed tree-top families in Whale Bay and Pirate
  Lagoon. Collision, interaction, sounds, RNG and guest actor loops are NOT
  reinstated. No second whole-guest scene draw is performed.
* Local keep-scenery, view distance and retention apply to that billboard
  subset. Settings are sampled once per workload. Only textures referenced by
  selected animation frames are copied, deduplicated across views/passes.
* Native rendering restores canonical RSP/RDP/vertex and deferred TMEM state.
  It uses a separate identity/scratch namespace and existing pack/mipmap
  resolution. Resources remain owned through workload/present retirement.
* Opt-in performance diagnostics separate guest, audio, hashing, copying,
  decoding and presentation costs, with recent interval percentiles and
  workload counts. No periodic GPU drain or per-frame diagnostic logging is
  introduced. Normal diagnostic collection stays off.

Still incomplete -- do not infer parity from the enabled scenery controls
-----------------------------------------------------------------------
* Native high-detail animated vehicle pose/shading/head/attachment construction.
* Broad level-mesh and non-billboard actor retention/draw-distance parity.
* Independent FOV, animated/effect ranges, persistent water detail and guest-
  side culling settings. They remain canonical in Experimental Rollback.
* Independent sound-category mixing remains deferred; master volume is local.
* A demonstrated steady 180 FPS for three/four-player races and cutscenes.
* Cross-revision visual qualification and translucent depth/order parity.

Existing preferences are preserved for single-player and legacy modes. Those
unsupported local controls remain disabled in an experimental session rather
than being made to write differing values into checkpointed simulation state.

Verification performed
----------------------
Windows/Linux compilation and static source/package checks only. The compiled
scenery, presentation-ownership, state-store and checked-memory fixtures were
NOT executed. No launcher, game, private probe, CTest or runtime packaging
self-test was run. Gameplay/GPU/network qualification is left to the user.
Compilation alone does not prove improved FPS, boss rewards, WAN stability or
absence of visual defects. Protected dependencies/generated sources were not
edited. See LOCAL-SCENERY-STATUS.txt for the detailed scope and evidence.

Suggested user check
--------------------
Use the same settings/FPS target on all test runs. Enable Graphics -> Camera
and scenery -> Experimental performance diagnostics before a three-player
session. In Whale Bay and Pirate Lagoon, compare tree tops, depth/fading and
seams; then compare frame pacing before/after the first racer finishes and
during cutscenes. Close normally and share host/client runtime.log files.
These measurements determine the remaining bottleneck without guessing or
lowering image quality. Also check a boss finish, menu transitions and a client
leaving; their working flow is intended to remain intact, not newly qualified.
