1.6.0-dev-98357ff

Linux builds failing CI, No Linux build for this release

1.6.0-dev-98357ff (2026-10-15 7:01:14 PM)

Round 7 of the audit re-walked the entire mod with fresh eyes - scaler/present
lifecycle, parser/translator, runner/compiler, tests/CI - and produced 25 new
findings (2 High, 11 Medium, 12 Low) plus a reopened MAINT-03. The Priority
Matrix's entire Immediate row shipped as one low-risk patch the same day,
closing both High findings and the unchecked-SDL-return cluster behind them;
the vendored-SDL3 build break (SDL >= 3.2 made SDL_UploadToGPUTexture return
void) and the CI clang-format gate were fixed on top of it. The WAV-fallback
diagnostics learned to name the failing asset and the MPQ archive it came
from. The ledger stands at 44 open findings (0 High / 14 Medium / 30 Low);
deltas.md remains the single record of closures.

Added

Audio

• DEVILUTIONX_WAV_FALLBACK CMake option: builds the SDL_LoadWAV_IO
  fallback path that recovers the WAV files SDL3_mixer 3.2.0's stricter chunk
  scanner rejects (zero-length or past-EOF data-chunk sizes in the tail).

Testing

• preset_smoke is registered with ctest against a committed good + malformed
  corpus (multi-pass/feedback/history/LUT presets; malformed presets must fail
  with a named error string, never a crash) - the round-5 fix that closed the
  audit's last High finding. It landed inside this release window and carries
  its own full entry above (1.6.0-dev-a67a4c0).

Packaging

• Packaging/windows/clear_gpu-scaler.log.cmd: one-click reset of
  gpu-scaler.log next to the game executable for Windows field reports.

Changed

Audit ledger

• audit.md refreshed by the round-7 sweep (+25 findings, MAINT-03 reopened),
  then re-scoped to the open set only (44 findings: 0 High / 14 Medium /
  30 Low) after the quick-fix batch; deltas.md gained the closure records
  for everything shipped in this window.

Removed

• .github/workflows/cache-cleanup.yml, .github/workflows/
clang-tidy-check.yml, .github/workflows/translations.yml - the MAINT-03
  zombie workflows, still pinned to the nonexistent actions/checkout@v7,
  actually deleted from the tree this time and verified with git ls-tree.
• .github/workflows/wav-fallback-test.yml - the Windows-only A/B workflow
  built to capture the fallback mismatch, retired once the diagnostics it
  drove had landed (added and removed inside this release window).

Fixed

GPU scaler (round-7 quick-fix batch)

• ROB-14 (High): activating the shader filter no longer destroys the
  SDL_Renderer while the global display texture still references it - the
  stale handle is dropped first, mirroring dx_cleanup, so toggling the
  filter can no longer hand a freed texture to the next
  ReinitializeTexture() (the virtual-gamepad half of the finding remains
  open as ROB-22 by design).
• ROB-15 (High): Runner::Run captures the present target BEFORE the
  double-buffer flip - feedback-driven final passes no longer present a
  permanently one-frame-stale image plus one undefined frame after every
  load/resize.
• ROB-16: a failed SDL_MapGPUTransferBuffer (device loss, allocation
  failure) now costs one skipped frame like every neighboring stage instead
  of memcpying to NULL outside the failure contract.
• ROB-17: SDL_BeginGPUCopyPass is null-checked (fail visibly + empty
  submit instead of blitting stale contents with a clean log). The
  upload-result half was reverted in an r7.6 fixup: vendored SDL3 (master,
  post-3.4.0) follows the SDL >= 3.2 API where command-encoded copy calls
  return void - per-call failure reporting was removed upstream, the bool
  capture broke the CI build, and the checked
  SDL_SubmitGPUCommandBuffer is the failure net.
• ROB-24: a failed Vulkan-retry window claim now nulls gWindow, so the
  Shutdown() path cannot release a window this device never claimed.
• ROB-25: an unexpected (non-custom) filter value is now counted and
  throttled-logged as a present-path skip instead of presenting a permanent
  silent black screen with a healthy heartbeat.

Audio / WAV fallback

• Load-failure diagnostics name the asset: SoundSample::SetChunk carries the
  asset path and - for MPQ-sourced assets - the archive path (exposed via
  mpq_reader's new path()), so a rejection logs
  Failed to decode audio file / <path> (<archive>) / <SDL_GetError> instead
  of a bare SDL error, alongside the 56-byte head hex dump from the previous
  round.

Parser

• MAINT-09: the #reference merge comment now states the real semantics
  (the LAST parent chain providing passes wins; textures/parameters
  accumulate).
• MAINT-19 (partial): texture validation names the actual problem (empty-name
  vs missing-path) instead of sending field debugging the wrong way, and the
  dead 4-argument ParsePreset declaration is gone. The bare shaderpath
  behavior decision remains open with the maintainer.

CI / build hygiene

• The CI build break with vendored SDL3 master (void value not ignored,
  SDL_UploadToGPUTexture bool capture) is fixed - see ROB-17.
• The clang-format check workflow is green for the first time against the
  mod: gpu_scaler.cpp had shipped with a whole-file indentation regression
  from the previous batch's tooling (whitespace-only, compiled fine); the
  canonical Source/.clang-format formatting is restored and the file's diff
  collapses back to the surgical change set, and 12 pre-existing violations in
  soundsample.cpp are fixed. Both files are token-identical to their
  pre-format state (verified by whitespace-stripped byte diff); all 698
  C/C++ files under the CI's check-paths now pass.

Full Changelog: gpu-shaders-v13...98357ff