Draw palette and port 0x31 changes from the scan line they land on

[bubio]

Stacked on #15.

vraminfo found the palette latched per scan line — 「PC88 のパレットの反映は hblank 毎(?)に行われる」 — and tried switching settings halfway down the display: 200 → 400 lines and graphics on → off both work (「可能なようです」). Port 0x32 digital → analog has no effect, and port 0x30's 80/40 columns and colour/mono garble the screen. We drew every frame with the registers as they stood at the frame boundary, so VRAMTEST T showed one flat colour where the page shows bands, and U used one mode for the whole screen. This is VRAMTEST #13.

Change

• Log: the bus records the registers the page saw take effect mid-frame (palette 0x54-0x5B, background 0x52, the analog background, and port 0x31's graphics on / colour / 200-line bits) with the CRTC scan line of each write that changes something (RasterVideo.swift). The log is kept per frame and handed over when the scan line counter wraps to 0 (CRTC.onFrameStart), so it does not depend on runFrame. Reset and state load clear it.
• Bands: RasterFrame.bands() turns the finished frame into bands of output rows. A display change takes effect from the next line, and CRTC lines are scaled to the 400-row picture (×2 at 15kHz). Retrace changes are moved ahead of the display changes, register by register. A program that sets the top in the retrace and changes it partway down (T, U, and any raster effect written the usual way) then gets the frame the hardware draws every time.
• Drawing: FrameCompositor draws each band with its values; the four renderers take the rows to write (default 0..<400). A frame with one state all the way down takes the old path with the old values.
• Ports 0x30 and 0x32 stay per frame. Reproducing the garbling is out of scope.

Results

• VRAMTEST T: 8 digital / 32 analog bands at 24kHz and 15kHz. U-1: the R plane in 400-line form below the switch. U-2: no graphics below it. U-0/3/4 unchanged.

• Tests: 917 pass (8 new in RasterVideoTests; the end-to-end one fails if bands are switched off). Release builds pass.

• Regression: identical to Let the V1S text DMA take CPU time while the display runs #15 (the TheHospital difference is the same 1297 px, everything else identical).

• 1,102 disks booted for 60 s in V2, Let the V1S text DMA take CPU time while the display runs #15 vs this branch: 1,092 identical and 10 different. All 10 are palette writes that land during the display and now split the frame:

  • palette animation: Crystal Prison, Tritorn, Sylpheed, Melt Down, Alpha DEMO, Seena, Scotty, Storm
  • a fade or flash caught mid-frame: Emerald Dragon, Navitune

  That is what the page describes. Whether the hardware also splits these frames can't be checked without one. If any of these titles is meant to write only in the retrace, this is where to look.

🤖 Generated with Claude Code

https://claude.ai/code/session_01EtarYVwQy1YVfWGChE5eg4