Denise: horizontal DIW comparator flip-flop model

[LinuxJedi]

What

Replaces the span-based horizontal display-window clip with the hardware comparator model (vAmiga Denise::updateBorderBuffer, hardware-verified): one window flip-flop, SET on an exact 9-bit HSTART counter match, CLEARED on an exact HSTOP match, evaluated per lores position with the register values current at that position. No match = no change, so degenerate windows never open or never close, and the state carries across lines and frames.

Hardware details modelled:

• The counter starts each line at the hblank edge (lores 0x24) and wraps 0x1C8 -> 2 near the line end, so positions 2..0x23 are a line's tail, not its head.
• OCS Denise does not reset the counter on lines 0-8: it free-runs modulo 0x200, letting otherwise-unreachable stop values (0x1C8..0x1FF, i.e. any OCS stop byte >= 0xC8) fire during vertical blank.
• Vertical sync leaves the flip-flop set.
• DIWSTRT/DIWSTOP writes reach the comparators one colour clock after the write cycle.

Consumers converted: background border split, a per-pixel window gate in the playfield painter (replacing the display_window_x span pair and its stop <= start -> +0x100 heuristics at that site), and a post-compositing mask that repaints closed intervals as border (keeps sprites out of closed regions; border sprites exempt). The painter's shifter-origin anchor stays paired with the DIWSTRT value whose comparator fired, so mid-line rewrites move where the window opens without moving the fetched picture.

Also fixes a palette bleed the diw tests exposed: a colour write in the horizontal-blank tail attributes to the previous output row, behind same-beam-line events that already seeded the next row's base palette - the written entry is now patched into that base. Previously a one-line COLOR00 marker set during hblank (vAmigaTS diw1 block 1) vanished entirely.

Test updates

Three unit tests encoded artifacts of the old span clip and now assert comparator behaviour:

• a later-HSTART rewrite no longer re-clips an already-open window (the flip-flop stays set),
• a window edge sits at the new HSTART's comparator match, not at the write position,
• an OCS stop byte >= 0xC8 never closes, so the BRDRBLNK test uses a closing stop.

New unit test covers cross-line carry (unreachable HSTART + late stop rewrite).

Results (vAmigaTS, % pixel divergence)

case	before	after
Agnus/DIW/OLDDIW/diw10	28.3	9.0
Agnus/DIW/OLDDIW/diw9	65.5	60.0
Agnus/DIW/OLDDIW/diw8	57.1	54.7
Agnus/DIW/OLDDIW/diw1	63.5	60.7
Agnus/DIW/OLDDIW/diw2	50.3	48.0
Denise/DIW/DIWH/diwtim0	5.5	4.9
Denise/DIW/DIWH/diwtim3b	16.3	14.6
Denise/DIW/DIWH/minmax	9.8	9.3

Agnus/DIW/DIWV improves slightly across the board (one +0.3 on onoff1).

The dominant residual in the OLDDIW bucket is not the window model: the whole bucket runs DDFSTRT $30, and Copperline places that early-DDF picture 2 lores pixels left of vAmiga (a uniform shift; with it factored out diw1 measures 7.9%). That is the Agnus/DDF fetch-placement class (oldhwstop* cases at 50-65%) and is tracked separately.

Verification

• cargo test --release: 1266 + 30 + 4 pass.
• cargo clippy / cargo fmt --check: clean.
• image_regression --ignored: 7 pass + the documented host-flaky ocs_bpu7_ham (fails on clean main too).
• Byte-identity vs main (raw screenshots): KS3.1/A1200 boot, Zool, Gen-X, Inside The Machine, Hamazing - all identical.