[phobos665]

Toolkit fixes from bringing up Constantine (5343000F, XDK 5849). With these it boots, plays its Bink intro movies, reaches the main menu and plays its first level at a native 60 fps, with Constantine and his apartment drawn. The title project itself stays local (titles/ is ignored).

Changes

• An immediate-mode draw at End runs the program the XDK started (c3bd8eba). Swap scales the 1280x480 supersampled screen down with one triangle: the XDK starts a two-instruction program (oPos = v0, oT0 = v9) through the push buffer and calls Begin / End. End drew it under the title's 2D program, which reads its texcoord from v2, so every pixel sampled one texel and each frame after the movies was a single flat colour. End now uses pushed_program_begin as the walk's own draws do.
• Visibility tests are answered from host occlusion queries (dd28b57d). They were answered "0 pixels". Constantine culls its scenery and its hero by them (53 a frame), so neither was drawn.
  • The walk now hands SET_ZPASS_PIXEL_COUNT_ENABLE (0x17CC) and GET_REPORT (0x17D0) to the HLE.
  • The RHI counts samples in between: new rhi_occlusion_*, one Vulkan query per draw, one D3D11 query per test.
  • GetVisibilityTestResult answers with the matching test's count, scaled to guest samples.
  • All closed tests are read back in one GPU wait. One at a time it was about 80 waits a frame.
  • RECOMP_HLE_D3D8_VISIBILITY=0 switches it off.
• Frame dumps and F11 see frames shown through the video overlay (39dbec03). They now catch movies presented by UpdateOverlay without a Swap.
• A thread that only polls the clock sleeps between polls. Constantine's Bink player checks whether the next frame is due (a mutex taken and released) and reads the TSC in a loop between movie frames: about two million kernel calls a second, 83% of a core, during every movie. A thread that reads the TSC 256 times within 1 ms, four windows running, now sleeps 1 ms with the guest lock dropped. RECOMP_TSC_POLL_SLEEP=0 turns it off. Over 50 s of intro movies, main-thread CPU went from 44.2 s to 6.3 s, with the same movie frames at the same times. In-level uncapped is unchanged.
• A switch's index walk stops at code something names (f61788a5, tools/disasm). The disassembler's walk over a switch's index bytes treated 56 8B (push esi; mov esi, ecx) as two more indices, because the switch has about 156 cases. It started a function in the middle of the mov, so the lifted method lost esi = ecx, its vtable calls went wild, and Select Difficulty showed no options. The walk now stops at a plausible function start that's named as code. Compared with the old disassembler, Constantine differs by exactly that one bogus entry. Other titles change only when re-lifted.
• Seeds for Constantine (259d78e7): six thread starts, plus a callback (0x001DF280) that the disassembler had run into from a jump table before it. It was called over 1000 times unresolved.
• 1e4536a8 is the same immediate-mode vertex fix as in Xbox Live without a console backup: generated console, Insignia registration, updater from the player's dashboard #108, cherry-picked here.

Measured (macOS arm64, Vulkan)

• In-level, capped: a steady 300 swaps per 5 s (60 fps).
• Uncapped: about 430 swaps per 5 s, against about 505 with RECOMP_HLE_D3D8_VISIBILITY=0. That run also draws less, because the culled objects aren't drawn.

Not done / needs checking

• The D3D11 half of rhi_occlusion_* has not been built or run. It needs a Windows build, plus the TS2 and BLiNX regression before merging, because the End change, the visibility tests and the clock back-off affect every title.
• Still open on Constantine:
  • 7 shadow-map draws read all-zero vertices;
  • 4 of 8 sound buffers are never heard.