DKR-R 1.0.5 Beta 15 Playtest 4 - bounded output-efficiency 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.

Date: 2026-10-03

IMPLEMENTED
1. Private audio RSP DMA now validates the entire RAM and DMEM transfer once
   before copying. Preserve the reviewed hardware 24-bit/8-byte normalization,
   inclusive lengths and byte-lane layout. Reject overflow, invalid invocation,
   DMEM overflow, RAM overflow and partial final byte-lane words before writes.
   Keep operation-budget checkpoints at DMA entry and generated RSP labels.
   Only NEW Patch Pipeline output changes; protected RSP/runtime inputs stay
   untouched. This removes per-byte scalar trap calls, not the safety boundary.
2. Compute the immutable render-image diagnostic hash on first request, with
   concurrent-reader once-only initialization. Ordinary frame publication no
   longer hashes an additional 16 MiB. Network checkpoint, journal and initial
   transfer hashes, immutable copies and GPU retirement gates are unchanged.
3. Query the actual complete draw-observation count before sizing its vector.
   Avoid zero-initializing the full 8192-record (~608 KiB) capacity every tick.
   Preserve all-or-nothing sidecar rejection and reject inconsistent capture.
4. Reuse the confirmed audio conversion buffer instead of allocating it for
   every block. Device output remains confirmed-only; stereo order, frequency,
   local master gain and output EQ are unchanged.
5. Explain and disable only unsupported live camera/scenery and sound-category
   controls during an active experimental match. Preserve saved preferences;
   normal single-player and legacy online controls are unchanged. Renderer,
   texture-pack, HUD, master-volume and output-EQ controls remain independent.

REVISED SCOPE / NOT CLAIMED FINISHED
The user chose performance first and explicitly deferred the independent scene-
construction and sound-category-mixer redesigns. The owned CPU continues using
canonical camera/scenery defaults and an authored category mix. No peer's local
preferences are imposed on another participant, or written into shared replay.
Full local scenery/category parity is deferred, NOT implemented by this build.

The earlier alAdpcmPull out-of-range guest-address failure has not recurred in
the latest supplied/local logs. Its precise bad operand and origin remain
unknown: do NOT call it fixed by the DMA optimization. Existing bad-address,
register, generated source/line and guest-stack diagnostics remain intact. If
it recurs, close the game and preserve both participants' newest runtime.log.

BASELINE / PERFORMANCE EVIDENCE
The accepted Playtest 3 ZIP, AppImage, hashes and pre-edit sources are preserved
outside the public packages in playtest3-local-parity-baseline-20261003.
The latest real user log contains a three-owner sample near 179 rendered FPS,
matching frame/confirmed counters, zero replay corrections and no rejected
packets in that sample. It is evidence about that session, NOT a WAN or every-
scene qualification. No FPS improvement is measured or claimed for this build;
the changes above remove source-identified redundant work without quality cuts.

UNCHANGED
Legacy backend algorithms; vanilla boot/menus and scene-epoch authorization;
save verification and online-only save/Pak storage; initial-state transfer;
texture-pack activation; 3/4-player viewport/music and HUD fixes; departure and
return-to-lobby flow. No direct edits to submodules, RecompiledFuncs or
RecompiledPatches. No gameplay shortcut, fake scene acknowledgement, pointer
masking recovery, skipped audio voice or default quality reduction.

VALIDATION / TEAM CHECKS
Compilation and static source/package/contract checks only. Regression cases
for DMA lanes/range rejection and concurrent lazy-image hashing are authored
and compiled, NOT executed. No gameplay windows, self-tests, CTest or automated
multiplayer tests are run, following the user's instruction.

Install the exact Playtest 4 packages on EVERY experimental participant; the
owned contract changes and rejects prior packages. Team testing must check:
- Initial launch and saves with two, three and four participants.
- Character/Track Select, previews, all race modes, Adventure bosses, finish,
  rewards/key consecutive cutscenes and return to menus.
- Normal audio, no static/pops/missing voices, local master volume and EQ.
- HUD, original/pack textures, interpolation, water and transition framing.
- Client exit, return to lobby, launch again and unchanged single-player saves.
Compilation is not proof of gameplay, DSP equivalence or Internet stability.
