Faces cut into loose piles collapse outward instead of draining column by column (t93)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

docs/DESIGN.md

@@ -555,6 +555,21 @@ above, including:
   keeps its chunk awake to retry. Burning fuel beds keep their tuned settling behavior.
   Grain stamps, sealed corners, transport barriers and per-chunk random streams still apply.
 
+
+- **Vacuum-cut loose faces spread while falling (t93 reopened, 2026-10-28).** The owner
+  reports that stone-only and mixed faces still stand after vacuuming into settled piles.
+  Sticky-step retries alone do not address the descending column formed by always choosing
+  straight down first. A falling grain on an exposed vertical face, backed by supported
+  powder, prefers an open outward diagonal. With at least 12 grains of backing above it
+  and unmoved support below that backing, it spills across up to eight clear grains and
+  falls up to eight clear grains. This is a local pressure approximation: the falling
+  outer column must drain quickly enough to let the body of the pile release. Shallow
+  faces keep the one-grain step, preserving small heaps and conveyor feeds. Check the
+  whole horizontal/downward path for occupancy, parts and cleats, including other chunks.
+  Keep one stamped move per tick and the existing eight-grain worker write halo.
+  Resting heaps do not enter this path; world generation keeps its recorded settling rules
+  and seed layouts.
+
 ## Ideas not built yet
 
 - **Concrete**, as a stronger wall material from rubble.

docs/PROJECT_MAP.md

@@ -254,10 +254,10 @@ Exploration only: never changes grains or the random stream.
 - functions: fogAlloc, fogMark, fogSeen, fogInit, fogLook, fogRevealAround, fogRevealAll
 - variables: FOG, FOG_SENSE, FOG_SENSE_FADE, fog, FW, FH
 
-### src/sim/cells.js (240 lines)
+### src/sim/cells.js (272 lines)
 Cell helpers (setCell, move, enter) and the falling-sand movement of powders, liquids and gases (simMove).
 - functions: setCell, swapCells, swapCellData, move, enter, enterFree, powder, passable, liquid, liquidSearch, gas, wakeBotPowder, simMove, moveChunk
-- variables: SORTP, SORTP_FUEL, LSTEP, LFILM, CORNER_SEAL
+- variables: SORTP, SORTP_FUEL, LSTEP, LFILM, CORNER_SEAL, POWDER_FAN
 
 ### src/sim/heat.js (266 lines)
 Heat: conduction, melting and freezing, reactions, fire and air, steel.
@@ -630,6 +630,7 @@ The frame loop and boot.
 - `test/harness.js`: Loads the simulation (the src/sim/ files listed in src/manifest.json, joined in order into one script, the same way build.py joins them) so scenarios can drive it headlessly in Node. Errors point at the real file and line (sim/pipes.js:42), not at the joined script.
 - `test/light-cases.js`: Lighting behaviour checks executed by the GPU browser harness.
 - `test/light-readback.js`: Read float source channels, including transmittance, from the WGSL light pass.
+- `test/pile-cut.js`: Shared Node/browser reproduction: vacuum the foot of settled 60-grain heaps, one stroke per tick.
 - `test/play.py`: Play-test in a real browser: mouse, keyboard and a mocked gamepad drive the game; save and continue. Screenshots go to shots/play/ (or the folder given as the first argument). Pixel look unless the second argument is 'smooth'.
 - `test/run_all.py`: Run the test suites and report.
 - `test/t_bar_stacking.js`: Resting bars snap into pairs of grains and tip off narrow supports.
@@ -668,6 +669,8 @@ The frame loop and boot.
 - `test/t_smooth.py`: The smooth look: the WebGPU pass compiles and draws, and the thermal lens still works.
 - `test/t_stone.js`: Uses for waste rock and sand: glass bricks from a fan furnace, the Saw, the silo and builders' orders.
 - `test/t_threads.js`: Shared allocation, checkerboard independence, and real worker-count determinism.
+- `test/t_vacuum_collapse.js`: Real nozzle cuts into settled heaps, inline and through the real Node worker pool.
+- `test/t_vacuum_collapse.py`: Vacuum cuts in the real pixel frame loop with active browser workers and hardware graphics.
 - `test/t_water.js`: Water conservation, settling, ledge pouring, boundaries, pipes and phase changes. Tall worlds keep geothermal heat out of the ponds.
 - `test/t_weather.js`: Weather: no more springs; steam rises into clouds and rains back; showers; snow when it is freezing.
 - `test/t_world.js`: Trees from acorns, charcoal, brick, density, water/lava/snow/ice, full-bag refunds, drill reach, ballistic blowing, snowy mountains.

docs/TESTING.md

@@ -78,6 +78,14 @@ checks waking within eight ticks separately from the probabilistic consumption o
   gamepad and world sizes. Screenshots go to
   `shots/play/`. Pixel look by default; `python3 test/play.py shots/play smooth` uses the smooth
   look.
+- `test/t_vacuum_collapse.py` cuts settled stone and mixed heaps with one normal-width
+  vacuum stroke per simulation tick in the real frame loop. It uses the shared
+  `test/pile-cut.js` fixture, requires actual helper dispatch (forced for this small world),
+  checks both faces within 180 ticks, and saves time-series pixel screenshots in `shots/`.
+  `test/t_vacuum_collapse.js` uses the same cuts inline and with Node helpers, including
+  bot-near/far, grain conservation, equal material bytes and forced-awake resting stability.
+  It also checks a broad 120-grain wall within 600 ticks and a spill stopped by cleats
+  across chunk edges, where the source chunk's one-grain barrier flag is clear.
 - `test/t_smooth.py` checks WebGPU smooth shapes, seams, joins, fog and the thermal lens.
 - `test/t_gpu_golden.py` compares 28 cropped frozen reference PNGs (384x288, under
   `test/fixtures/golden/`) against WebGPU: renderer, joins, sky and four light scenes,

docs/agent-log.md

@@ -5,6 +5,59 @@ 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 reopened, vacuum-cut faces drain in columns (Codex)
+
+- Read AGENTS, the t93 task/log and 807217b; inspected both new owner screenshots.
+  The remaining reproduced cause is straight-down movement taking priority over outward
+  fall: an exposed edge stays a descending column, blocking the next column's diagonal
+  exits. This is a movement/scan-order bottleneck, not lost wakes in this reproduction.
+  `setCell` wakes the cut; worker wakes merge each checkerboard pass. A radius-6 vacuum,
+  one call per tick into a settled 60-grain RSTONE heap, leaves a 39-grain cliff; after
+  120 ticks it is still 18, with exactly the same series when all chunks are forced awake.
+  RSTONE already has STICK=1. Inline and actual Node helpers agree, bot near or far.
+- Added outward downward movement only for a falling exposed vertical face backed by
+  supported powder. Retain one downward step per tick, occupancy, parts, barriers and
+  burning-grain guard. Resting heaps do not enter this path. `POWDER_FAN` goes to workers
+  and is disabled during the existing legacy worldgen settle to preserve all seed hashes.
+  Rejected a horizontal-step experiment (forced-awake creep) and a broad falling-edge
+  rule (spread free-falling blocks too far); no old test expectation was changed.
+- New shared `test/pile-cut.js`: two 60-grain triangular heaps, 600 settling ticks,
+  stone and stone/copper ore/iron ore/dirt, then actual radius-6 vacuum strokes.
+  New Node suite passes 40 checks: cuts inline/2 helpers, near/far bot, conservation,
+  exact inline/helper material bytes, initially stable heaps and 600 forced-awake
+  stable ticks afterward. Browser suite uses isolated HTTP, real NVIDIA Blackwell GPU,
+  real frame loop and two helpers with confirmed dispatch (small fixture forces dispatch).
+  Verdict is recorded at exactly tick 180. Control with face spreading disabled:
+  cliff stone/mixed 33/32 at cut end, 14/13 at 120, 8/5 at 180; browser assertion fails.
+  Fixed browser reaches <=2 by 120 and passes at 180. Earlier Node red run fails eight
+  collapse checks (its mixed fixture included clay); existing clay collapse checks pass.
+- Full `python3 test/run_all.py --browser`: ALL PASSED, 34 headless + 9 browser suites,
+  including collapse 17, threads 12, vacuum 40, gameplay 361 and browser vacuum 1.
+  Log: `/tmp/t93b-all-browser.log`. No existing test was weakened. Browser regression
+  is included in the suite runner. Original seed hashes and liquid-sinking checks pass.
+- Saved and inspected `shots/t93b-before.png` (control at cut end) and
+  `shots/t93b-after.png` (fixed after 180 ticks): vertical cut faces become slopes.
+  Fixed series: `shots/t93b-time-{0,60,120,180}.png`; matching control series:
+  `shots/t93b-control-time-{0,60,120,180}.png`. Browser control log:
+  `/tmp/t93b-browser-matching-control.log`. Small initial browser passes stayed inline
+  despite a pool existing; final regression requires actual `simPool.activeMax > 0`.
+- Requested idle profile, real GPU, no concurrent tests:
+  `python3 tools/perf.py --scenes settled,heat --looks pixel --zooms 6 --compare tools/perf-after-t88.json`.
+  Rule-disabled control -> final total ms/tick: settled 1.9855 -> 2.0226 (+1.9%),
+  heat 2.7027 -> 2.7734 (+2.6%). Against historical t88: +12.0%/+5.7%.
+  Movement ms/tick 0.9639 -> 0.9818 / 1.3481 -> 1.3885; unchanged heat work also
+  varies 0.9412 -> 0.9593 / 1.2724 -> 1.3019. End awake counts identical, 425/494.
+  These timings show small run-to-run differences, not exact equality; historical settled
+  time remains over the profiler's 10% comparison threshold. Evidence:
+  `/tmp/t93b-idle-control.json`, `/tmp/t93b-idle-final.json`, `/tmp/t93b-idle-summary.log`.
+  An earlier control profile overlapped headless checks and was excluded from this comparison.
+- Files: `src/sim/cells.js`, `src/sim/workers.js`, `src/sim/worldgen.js`,
+  `test/pile-cut.js`, `test/t_vacuum_collapse.js`, `test/t_vacuum_collapse.py`,
+  `test/run_all.py`, `docs/DESIGN.md`, `docs/TESTING.md`, generated project map,
+  t93 task and this log. No commit. Owner playtesting of their larger worlds remains open.
+
+
 ## 2026-10-28: t92 keep pass-through, restore original loose-grain behaviour (Codex)
 
 - Read the 7293787 then dadf62a t92 history and original drill in 90f2d3f. Restored the