erik/acdream
1
0
Fork 0
Code Issues Pull requests Projects Releases 6 Packages Wiki Activity Actions

docs(issues): #464 geometry round — punches vertex-identical to retail, 12 of 16 rejected by the ±12 rule in both clients; no authored wall above the arch

Browse source 
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
This commit is contained in:
Erik 2026-09-03 22:45:07 +02:00
parent 3a6aa7eef4
commit 2f7f45c2dd
1 changed files with 2 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

2
docs/ISSUES.md
View file

@ -115,6 +115,8 @@ then `/teleloc F4180113 39.502899 19.589222 181.119247 1 0 0 0`, zoom out, tilt
**Gate replay (lead, 2026-09-03 22:29, scratch worktree `s4-c1-gate` at the S4-c1 fix-round commit `2bd353234` with the capture copied in and a fifth `[Fact]` added):** `AlphaDepthTranscript_CathedralStairArch_MatchesRetailFrame2` **PASSES** — acdream's PM (mode, counterBefore) and PC (ov, counter, forceClear) sequence at the artifact pose is EXACTLY retail's: the sixteen building-portal punches, the eight exit seals, the counter. So the divergence is NOT which punches/seals happen or their order; it is their GEOMETRY on screen (what each punch fan covers, what each seal writes) or the building shell's depth at that face — the transcript gate cannot see either. AD-17 records the one known geometric deviation exactly there: a >8-edge or multi-polygon OUTSIDE view (the stairwell's open face IS multi-polygon: three exit polys of `0x114` + two of `0x113`) yields a zero-plane slice and that view's punch fans draw UNCLIPPED, while retail CPU-clips every portal polygon exactly. Next: compare the punch-fan geometry per portal against retail's clipped polygons (retail's `DrawPortalPolyInternal` clips the polygon against the current portal view's planes before drawing; a vertex-dump capture at the pose, or the replay harness's own clipped polygons, gives the retail-side outline) and check that the cathedral's shell is drawn with depth before the punches. (Side finding of the same replay: `AlphaDepthTranscript_TerraceEdge` FAILS at `2bd353234` — see the S4-c1 fix-round record.)
**Geometry round (lead, 2026-09-03 22:36–22:45):** a new retail template (`oh-capture-pm-verts.cdb.template`, `cathedral-stair-arch.pmverts.log`, 3 frames) prints every `DrawPortalPolyInternal` attempt's vertices; the scratch replay (`s4-c1-gate`, a `PunchGeometryDump` fact over the same harness) prints acdream's. Result: 16 attempts per frame on both sides, VERTEX-IDENTICAL (e.g. retail poly `12980fd4` = acdream building `0xF418000C` view 1: `(4.5018, 11.924, 21.8948)` …), and applying retail's own reject rule (every vertex on ONE of the x=±12 / y=±12 planes) to retail's list rejects 12 of the 16 — exactly the 12 acdream drops between collect and replay (`COLLECT punches=16 REPLAY fans=4`). The rejected 12 are the five of building `0xF418000B` and the seven of the HALL building `0xF418000A`, all on local y=12.000; the four drawn are `0xF418000C`'s at y=11.924. **So the hall's doorways are never punched in either client, and the punch geometry is not the divergence either.** The DAT then rules out a missing wall: the tower shell `0x01001FB7` (local y −10.165 … +10.165, z 0 … 53.7; 363 render polys) has NO polygon with a ±y normal on its far face above z 21.9 (only `poly 0x000B` at z 20–21.9), and the hall shell `0x01001FB3`'s +y face at local y=12.0 (the same world plane y=24.0) has none above z 21.9 either; the hall cell `0x109`'s near face on that plane is itself an EXIT portal at local y=+12. Nothing authored stands in the region above the arch. **Therefore the solid wall retail shows there is produced by the outside-view block's depth/colour ORDER — what is on screen in that region when the exit views are sealed and the interior is repainted (the gated Z clear, the `FlushAlphaList`/stamp, the seals, the repaint), not by a polygon we fail to draw.** Both event streams (PM/PC) match retail exactly, so the difference is in what each event DOES to the framebuffer. Next: pixel attribution at the tilted pose (which draws touch the region above the arch, in order, with their depth states) — a scripted tilt (the `mouselook` automation verb, in flight) plus a RenderDoc or recording-device capture; then the retail-faithful analysis of the diverging step.
## #463 — Soak `turn` samples can report a 0.1 ms GPU p50 (the turn window catches non-world frames)
**Status:** OPEN — found 2026-09-03 by the Campaign OVERHAUL v2 S3 review's perf re-measure.