Loose walls slump to a natural slope instead of eroding layer by layer (t93)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-28 18:58:40 -04:00
parent dadf62a28d
commit 807217b22c
7 changed files with 146 additions and 3 deletions
@@ -5,6 +5,50 @@ note or request it was for, what changed, the files touched, the tests run, and
 open or not verified. Keep each entry short; the git history has the details.


+## 2026-10-28: t93 loose faces slump without sticky-step stalls (Codex)
+
+- Inspected the owner's screenshot and traced `powder`, `enterFree`, `move`, `moveChunk`,
+  waking and checkerboard worker passes. Movement-only 40x20 walls on a flat floor, with
+  normal sleep countdown and the bot far away: sand reaches a <=1-grain surface step in
+  44 ticks; rubble (stone/copper ore/iron ore/dirt) in 89; 20% clay never reaches it through
+  2400 ticks (2-grain ledge). Always-awake control reaches it in 249 ticks, pure clay in 585;
+  STICK=1 takes 44 for every mix. Cause: probabilistic sticky diagonal steps throttle face
+  release, and three-tick chunk sleep can strand a legal step after missed rolls. Bottom-up
+  scan and stamps still allow sand to collapse; empty destinations are accepted before the
+  target-stamp check. No scheduler or stamp changes were needed.
+- Added a fallback only after a missed sticky roll: an open diagonal with another empty cell
+  below releases immediately; a legal shallow step wakes its chunk to retry. Retained the
+  original fast path, density/displacement rolls, built-surface sliding, sealed corners,
+  occupancy, transport and cleat checks, one-step stamps and per-chunk streams. Burning
+  grains/supports retain their tuned behavior: extending the fallback into burning beds
+  failed vent-fed furnace cleanliness. A shared-loop rewrite was also rejected after profiling.
+- New `t_pile_collapse` has 17 checks: all 800 grains retained; sand, mixed rubble and
+  20% clay faces <=2-grain surface steps by 180 ticks at offsets 0 and 17; cohesive pure
+  clay <=3. After 1200 forced-awake settling ticks, another 600 leave all heap material bytes
+  unchanged and chunks sleep afterward. Settled clay is 27 high versus sand 23. Exact final
+  test against original cells.js fails three collapse checks; final code passes. Added a
+  clay/rubble face spanning chunk rows to the real 0/1/2/8-worker/reversed-pass hash fixture.
+- Final `python3 test/run_all.py --browser` exited 0, ALL PASSED: 33 headless suites,
+  including furnace 9, collapse 17 and threads 12; browser gameplay 361, smooth 28,
+  console 18, caps 7, GPU 15, goldens 28, light 71 and Firefox 2. Evidence:
+  `/tmp/t93-final-browser.log`. Earlier full browser run also passed. Diff whitespace check passes.
+- Final isolated hardware profile, NVIDIA Blackwell, requested command:
+  `python3 tools/perf.py --scenes settled,heat --looks pixel --zooms 6 --compare tools/perf-after-t88.json`.
+  Original-code control -> final total ms/tick: settled 2.0179 -> 1.9790 (-1.9%),
+  heat 2.7255 -> 2.7650 (+1.5%). Against historical t88: 1.8056 -> 1.9790 (+9.6%),
+  2.6231 -> 2.7650 (+5.4%). Final movement costs 0.9715/1.3816 ms/tick versus control
+  0.9951/1.3647. Earlier profiles varied, including unchanged heat work; final run had no
+  concurrent tests. Evidence: `/tmp/t93-perf-fast.json`, `/tmp/t93-idle-control.json`,
+  `/tmp/t93-red-exact.log`, `/tmp/t93-collapse-final.log`.
+- Captured and inspected pixel `shots/t93-before.png` and final `shots/t93-after.png`:
+  three 40x20 walls become sloped heaps after 180 full simulation ticks. Reproduce with
+  `tools/scenes/t93-pile-collapse.js`, `--ticks 0` / `--ticks 180`, `--zoom 3 --no-light --freeze`.
+  No page errors. Rendering code unchanged; smooth screenshot not taken.
+- Files: `src/sim/cells.js`, `test/t_pile_collapse.js`, `test/t_threads.js`,
+  `tools/scenes/t93-pile-collapse.js`, `docs/DESIGN.md`, generated `docs/PROJECT_MAP.md`,
+  t93 records and this log. All subtasks ticked; t93 moved to Review with a completion
+  report and game checks. No commit. Owner playtesting remains open.
+
 ## 2026-10-28: t92 follow-up, visible dust and a botward channel through rubble (Codex)

 - Traced shake events through `drainEvents` to `spawnDebris` and `drawParticles`. Events were