Review: preserve live DMA2 reads and restore word-cost timing

[Alexbeav]

Summary This review branch combines the corrected GPU DMA2 linked-list state machine from #13 with the lower transfer cost demonstrated by RetroPortingToolKit#299 and resubmitted as RetroPortingToolKit#302. The change keeps DMA slicing. It reads each payload word at its own guest-cycle event, rejects stale widescreen-prepass metadata, consumes every elapsed event when service is late, and removes the obsolete polygon filter. It then charges one guest clock per transferred 32-bit word, including each node header, with no separate setup charge. ## Why PR RetroPortingToolKit#277 fixed Vampire Hunter D and Spot Goes to Hollywood by allowing the guest to change later ordering-table data while DMA remains active. Its 8-clock header and 5-clock setup charges also made large Final Fantasy VII ordering tables take about three times the prior runtime cost. FF7 then reached its stuck-channel path, canceled active walks, and lost battle geometry. The one-clock model is a validated compatibility model. This branch does not claim that it is a complete physical-bus arbitration model. ## Scope gate - Upstream base: 22fbbfcab962b25e010b1b8e94cb2fa140612f63 - Review head: 9dfb4174adea419c259b132eda512525cf9b8dea - Commit count: 4 - Changed-file count: 13 Commits: 1. f424a814 preserves live payload-word timing and prepass guards. 2. 99f775c8 closes the late-service, live-link, and polygon-filter review gaps. 3. bc88c06c restores one clock per transferred word.

4. 9dfb4174 versions the larger DMA save-state wire and corrects the test record. Changed files: - docs/internal/FAITHFUL_TIMING_PLAN.md - runtime/CMakeLists.txt

• runtime/include/boot_state.h - runtime/include/dma_gpu_ll.h - runtime/include/gpu.h - runtime/include/ws_prepass_guard.h - runtime/src/dma.c - runtime/src/dma_gpu_ll.c - runtime/src/gpu.c - runtime/src/main.cpp - runtime/tests/test_dma_gpu_linked_list_timing.c
• runtime/tests/test_savestate_status_protocol.py - runtime/tests/test_ws_prepass_guard.c The diagnostic counters and synchronous A/B switch from dma: charge DMA2 linked-list walks what the validated model charged RetroPortingToolKit/psxrecomp#302 are intentionally not part of this review. They can receive separate review after the transfer behavior settles. ## Source tests - The runtime and oracle targets build with Clang 22.1.8. - dma_gpu_linked_list_timing_test passes 20 consecutive runs. - ws_prepass_guard_test passes 20 consecutive runs.
• savestate_status_protocol_test passes 20 consecutive runs. - 61 of 62 enabled runtime tests pass. - sio_dualshock_rumble_test and spu_fidelity_test remain disabled upstream. - mod_runtime_test crashes on this branch and unchanged base 22fbbfca. - git diff --check passes. ## Remaining gate The exact review head still needs visible Vampire Hunter D, Spot Goes to Hollywood, and Final Fantasy VII routes before an upstream submission or update. ## Review focus Please review the zero-cost setup transition, the one-event-per-payload-word contract, late-service behavior, save-state fields, determinism data, and widescreen cache invalidation. Developed with AI assistance; validated as described (test evidence in PR body). AI writes the code and the PR, but I always test before I send something up. Happy to iterate on this process with your feedback. --- ## Summary by cubic Restores the one-clock-per-word GPU DMA2 timing model while keeping each payload word read at its own guest-cycle event, so guest rewrites during an active transfer still affect later words. - Charges one guest clock per transferred 32-bit word (header and payload) with no separate setup charge; the old model charged 8 clock cycles per header plus 5 per setup. - Removes the obsolete opt-in polygon-drop filter and its PSX_ND_SIB_FLAP_LAST / PSX_ND_OT_OPAQUE_LAST env vars. - Widescreen prepass now fingerprints cached nodes and commands; any live RAM mismatch invalidates the cached transform metadata. - Late service calls now consume every elapsed DMA boundary, not just one. - Bumps the save-state format to v6 to serialize payload_index; older save states are rejected at the header before any state is applied. - Adds unit tests for the prepass fingerprint guard and the v6 save-state wire. ^{Written for commit 9dfb417. Summary will update on new commits.}

[cubic-dev-ai]

All reported issues were addressed across 11 files

<sub>Reply with feedback, questions, or to request a fix.

Re-trigger cubic</sub>

[Alexbeav]

@cubic-dev-ai review this PR

[cubic-dev-ai]

@cubic-dev-ai review this PR

@Alexbeav I have started the AI code review. It will take a few minutes to complete.

[cubic-dev-ai]

No issues found across 13 files

_{Re-trigger cubic}