Crossing PASS (248 m, 26.6 s), TilePostReachProof PASS, DetSinCos PASS, canonical export identical with -voxel.CrtFma3=0 (86A99075), walk Connected 42,870 / 17,413. FAILED: one S/M/S/M pair on the owner's staged map shows worker generation 47.7 s vs 30.7 s static (+55%) and 83.1 s vs 60.9 s moving (+36%), request p95 +19-28%. The room-ownership containment test now runs per sample in the generic tunnel SDF, the core evaluator and both native fallbacks. Next round: same geometry (export hash must stay 86A99075), cost back to the tip's level.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ
Records the owner-playtest findings (stale editor build, owner assets vs harness defaults, the 0.1 slope freeze, ribs with every terrace setting at 0), the proven rib cause and its fix on fix/tunnel-shape, the reverted Catmull-Rom attempt and why its evidence was wrong, the in-editor ablation procedure (cvar read once at first Play), current canonical numbers, and the queue awaiting the owner.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MmSoF64BbRFmk9C3p1VsBJ
Bet A: the per-sample lookup it would remove is 5.1-6.1% exclusive, and feature-major visiting would touch 5.2-6.6x more lattice points than the current code evaluates; the swept kernel is branch-divergent. Killed on its own numbers. Bet B: 20.04% of static LOD0 density calls are exact duplicates with 80.6% donors already finished (~2.96 s of 18.3 s), moving 20.70% / 63.3%; re-derived independently from the raw trace. Its rung-2 prototype is committed unvalidated on sol/tile-sample-reuse after the Codex usage limit ended the round mid-rebuild; resume brief in the scratchpad.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TaFdXhTdYMkW1xrV6ApT6Y
- WORK-NEXT: state after 2026-09-16 -- Sol's review, the sol/deterministic-math result, new
canonical export/worms/field hashes, and the next branches in order.
- CODEMAP 3.2: VoxelMath::DetSinCos / IsFiniteFast rows.
- ARCHITECTURE: passage landings have had no connector to the (0,0) room since the owner's
d97373c playtest correction; the doc still claimed one. Records the open gap Sol found (the
random-reach fallback when the footing query declines is unproven) and the proposed fix.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TaFdXhTdYMkW1xrV6ApT6Y
Records why the merge must not be re-proposed without a new idea. WorldControlPoints keep the
authored room-floor anchors and walkable routes before cave warp; replacing them lost 1,146 fit /
975 walk-reachable cells, and the exact shared-segment variant was slower. The remaining large
costs are the geometry itself, so the exact-skip perf push has reached diminishing returns; the
reach proofs are now guarded by an automation test (8ef1a3b).
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TaFdXhTdYMkW1xrV6ApT6Y
Records the playtest-driven fixes (strate crossing, sealed-rock proof, option B floors, gap
D-floors), the cold-audit correctness round (including the tile-cache determinism bug), the perf
gains measured against committed-tip DLLs (static worker ~184 s -> ~28 s over the whole push, same
world), the current remove-one cost map, and the parked exact-cull branch.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TaFdXhTdYMkW1xrV6ApT6Y
Records in the repo what two days of perf rounds established: the in-game worker CPU trajectory
(~184 s -> ~34 s static), the remove-one ablation costs, the three approaches that were tried and
not kept (with where the parked work lives), the measurement lessons that each cost a round, and
the next candidates.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TaFdXhTdYMkW1xrV6ApT6Y
Measurement round (Codex; checked against its report):
- Per-tile identity for ab4bbe3's always-on LOD0 changes: a fresh 841-tile HEAD trace matches
traces recorded before them (MemoStaticOn/Off), keyed by (x,y,z,level) with
verdict/empty/triangles compared. 0 differences, 0 missing; LOD totals identical, 669,834 in all.
- The "missing 10 s" is not the spatial index: all six TileCacheWindow x SpatialIndex
combinations are within ~1% for a given window setting. TileCacheWindow=0 at HEAD is not
2990009's path. It still thrashes at L3/L4 (p95 ~250 ms). The default beats 2990009 on every
metric: static worker 48 -> 34.4 s, moving 125 -> 101 s, LOD0 generation p50 73 -> 66 ms
static and 70 -> 62 ms moving.
- voxel.SpatialIndex default -1 -> 1: forced-on was marginally best in every pair (within noise).
- The sampler's "unknown" ~13.8% LOD0 leaf is ucrtbase.dll!_finite (RVA 0x3C080), the non-inlined
CRT call behind FMath::IsFinite, not Core.dll. Named by parsing Core's export table (no export at
or below 0xC080) and aggregating callers: the tunnel chain, passage landing/floor checks,
GetDensityWithParams, GetDensityAt, the spatial index range and EvaluateSDFCached.
Gates: export OBJ unchanged (b3e5f4c3..9b377), capability 20,830 / 10,909, determinism assertion
passed.
Next: an inline bit-test IsFinite (the same predicate for every input, so field-exact).
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TaFdXhTdYMkW1xrV6ApT6Y
cfad3d1's message blamed cross-round drift on the owner using his PC. That was my unverified
guess, and it is false: he had not used the PC all day. The gap is systematic. The switch-off
baseline measured 58 s static worker CPU in two rounds hours apart, against 48 s at 2990009, and
LOD0 generation p50 is ~4% slower than at 2990009. Something left on with
voxel.TileCacheWindow=0 costs ~10 s. Queued: attribute it with in-tree switches (first suspect:
voxel.SpatialIndex auto-on at fused LOD1-2), never by running old commits.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TaFdXhTdYMkW1xrV6ApT6Y
Owner-approved, on the A/B measured in 83a0e4d: the classifier plus exact validation cost more
than it saved on the game path. With it off, ready p95 is 182 vs 190-195 ms, generation p95 is
125-127 vs 139-141 ms, and density calls drop 18%.
- GVoxelForgeOuterClassifierMode defaults to 0. Mode 1 stays reachable via
voxel.OuterClassifierMode=1 as an A/B instrument; the clamp is now 0..1.
- Mode 2 (validation samples reused as the mesher core grid) is deleted: the reuse grid,
FVoxelMesherCoreDensityGrid, GenerateMesh's ReusableCoreGrid parameter, bValidationCoreReused,
TotalValidationCoreReuseTiles and its log field. It could not fire: it reused the grid only
when validation disproved a uniform verdict, which never happened, and Mixed tiles never
validate. The mode-1 validation loop exits early again on the first disagreeing sample.
- All three GenerateMesh call sites checked; the density-count pointer binds after BandVoxHi.
Verified by Codex, checked by me against the logs:
- build ok; the measured process loaded the staged DLL, SHA-256 1024D8CF..04D
- export OBJ SHA-256 unchanged: b3e5f4c3..9b377
- a run WITHOUT the override prints outer_classifier_mode=0 validation_density_calls=0,
proving the new default took effect
- 841/841 tiles have identical triangle counts between mode 0 and mode 1. The 394 tiles mode 1
skipped as uniform produce 0 triangles in mode 0, so the strate band cut (applied by the
mesher, not by validation) was not hiding geometry.
Docs: ARCHITECTURE 8.10, CODEMAP (GenerateTileResult, ClassifyTile) and fable-idea T1.d record
that streaming no longer classifies by default. WORK-NEXT item 1 records the owner's playtest:
he can no longer fall through the world.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VZS94UH8bSsBeg5sZaxe2Y
Three-mode outer-classifier A/B on the headless game path, nested refinement off in all
modes, clean and diagnostics off. Runs made by the forked Codex session; the table was
assembled and checked by me from its logs (Saved/OuterABMode*_20260912.log,
OuterABMove*Headless_20260912.log).
static ready p95 mode0 182.4-182.7 mode1 189.7-194.8 mode2 190.0-194.7 ms
gen p95 mode0 124.9-127.0 mode1 139.1-140.6 mode2 140.3-141.3 ms
worker CPU mode0 184.9-193.5 mode1 201.2-202.3 mode2 201.2-201.3 s
density mode0 18.05M mode1 21.89M mode2 21.89M
moving ready p95 mode0 198.9 mode1 208.0 mode2 210.7 ms
density mode0 57.7M mode1 69.6M mode2 69.6M
The outer classifier costs more than it saves: with it removed (mode 0), every latency
figure improves and density calls drop by about 18%. Mode 0 also cannot remove geometry,
because it skips nothing.
Mode 2 (reuse the validation samples as the mesher's core grid) recorded 0 reuse tiles in
every run. That is by design, not a bug: the grid is only reused when validation DISPROVES
a uniform verdict, and it disproved none. Mixed tiles, which are the ones that need a
grid, never run validation at all.
Recommendation recorded in WORK-NEXT.md, pending the owner: default the outer classifier
mode to 0. The code default is NOT changed in this commit.
CORRECTION to f204413, whose message says the nested classifier's 3.5x could not be
reproduced. That was my error. I passed the streaming console variable
-voxel.UseBlockEarlyOut=1 to the export commandlet, which reads its own -blockearlyout=
argument (VoxelForgeExploreCommandlet.cpp:381, :1093), so both of my runs had nested
refinement off. Re-run with the right flag:
-blockearlyout=0 0.553 / 0.572 s
-blockearlyout=1 2.036 / 2.060 s 3.6x, geometry SHA-256 b3e5f4c3..9b377 identical
The 3.5x was real. Turning nested refinement off in streaming (f204413) was worth it,
and Sol's reading of the table was right where mine was wrong.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VZS94UH8bSsBeg5sZaxe2Y
Written by the classifier-fix Codex round before it was closed. The classifier is a
second evaluator -- interval proofs, exact certificates and final-field validation --
beside the canonical lowered one, and the runtime can validate the same tile again
before meshing. Any classifier kept long-term must consume the lowered representation
and reusable density data.
That round's other uncommitted edit was NOT kept: it flipped
GVoxelForgeOuterClassifierMode 1 -> 0, turning the outer classifier off by default.
That was A/B setup for an unfinished experiment, not a measured decision, so the
committed default (1, classifier on) is restored.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01X9DLqyXrJxG4TojzDrSmVw
Two WORK-NEXT items, one null result and one confirmed removal. Item 4 untouched.
ITEM 3 -- CANCEL OBSOLETE WORK AND PRIORITISE THE FLOOR: NULL RESULT
Clean -nullrhi, diagnostics off, 800 cm/s, 200 m of travel:
before after
request->ready p95 0.369117 s 0.367612 s
queue wait p95 0.066949 s 0.064483 s
generation p95 0.287926 s 0.283642 s
p95 travel distance 2.952938 m 2.940894 m
obsolete aborts 0 12
Flat. Aborting obsolete tiles, sorting from the true pawn position and prioritising the
occupied/next tile changed request-to-ready by 0.4%.
⭐ AND THE NULL RESULT IS ITSELF THE FINDING. Look at the split: queue wait is 65 ms of a
368 ms p95; generation is 284 ms. **The queue was never the problem, so reordering it
cannot help.** Only 12 tiles in a 200 m run were obsolete enough to abort. The floor is
late because generation is slow, not because work is badly ordered -- item 3 establishes
that by exclusion, which is worth more than the 0.4%.
The machinery is kept: it is correct, it costs nothing, and it will matter once
generation is fast enough for ordering to be the binding constraint.
ITEM 5 -- VERDICT CACHE REMOVED. It was pure overhead.
Clean moving-session audit:
outer calls/probes/hits 2,109 / 2,109 / 0
nested calls/probes/hits 19,904 / 19,904 / 0
enclosing hits 0
stores 15,877, evictions 15,365
⇒ It stored 15,877 proofs, evicted 15,365 of them, and hit ZERO times. A moving player
visits unique regions; the cache was filling and thrashing a 512-entry table under a
global lock for no benefit. Same-binary A/B: request p95 0.367645 vs 0.365978,
generation p95 0.285559 vs 0.285992 -- neutral within noise, so its cost was the lock and
the linear scans rather than time we can see.
Removed: the global cache, its lock, the entry table, the invalidation path, the
cache-only profiler fields and the synthetic probe. The proof classifier stays.
⚠️ I asked for this cache two days ago on the strength of a commandlet measurement that
re-visited the same region repeatedly. That access pattern is nothing like a moving
player, and I should have said so when specifying it.
VERIFIED BY ME: build 0 errors; export 0.563 / 0.565 / 0.566 s (35de59b was 0.545-0.563,
flat as expected); geometry SHA-256 b3e5f4c3..9b377 UNCHANGED; capability 20,830 /
10,909, connected.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012aheo2crrZmHuYtPpdY2Cb
Ordered by importance then size. Sol (read-only, xhigh) audited experimental @ aa62971
against main and delivered the two-sentence diagnosis this queue is built on:
"The operator abstraction leaked into the innermost loop, and streaming treats
'work submitted/applied' as 'safe to stand on.'"
Two corrections it made to us, both recorded in the doc:
- The fall-through is NOT a removed correctness barrier. main has no spawn gate
either, marks tiles loaded before applying, and collides only LOD0. It was fast
enough to hide the race. The owner's evidence was real; the inference was not.
- The loop inversion did not recover the 8x because it traded PER-SAMPLE
short-circuiting (~70% of deep-solid samples skipped the cave suite) for per-block
pruning. The fix is lowering the operator graph into a fused evaluator, not more
dispatch removal.
It also found that RMC cooks collision ASYNCHRONOUSLY and we never record
cook-complete -- apply_s ends at submission, so nothing we have ever measured says
when physics could actually collide.
Reading Plugins/RealtimeMeshComponent is now permitted for item 1, READ ONLY, by the
owner's decision.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012aheo2crrZmHuYtPpdY2Cb
Fr0zkaandClaude Opus 5