Files
VoxelForge/Standalone/Research/WorldAssemblerM3/Linux123/REPORT.md
T
Fr0zkaandClaude Opus 5.5 5e7e0eb1c1 research: region completion gate (L123, no production change)
The 18 ms gate is one game-thread tick, not compile cost (0.25 ms). MPSC handoff keeps ownership but still
polls per tick; velocity look-ahead didn't help (near -0.8 pt, p99 11.77 > 11.00). Both kept as a patch.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016dhxM43dnWvUu2B5Ph1iG1
2026-10-30 09:55:45 +01:00

10 lines
1.7 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
# L123: region completion gate
Five interleaved 30 m/s UE pairs show the gate is a game-thread poll delay. Core compile wall time is p50 0.251 ms (worker queue plus task p50 0.475 ms); request-to-observe is p50 18.89 ms, p95 21.68 ms. In 95% of traced completions, the next scheduler pass drains the MPSC result and resumes admission in that same pass; a small tail waits 2–4 scheduler frames. This is one tick interval, not separate compile, observe, and admit ticks.
The worker can safely hand off an immutable result through MPSC with its runtime state and `GenerationEpoch`. The game thread validates both before updating state and admitting tiles. Direct worker-side admission would touch game-thread state unsafely. This queue handoff preserves ownership but still polls once per tick; a game-thread wakeup/continuation is the safe way to try removing that delay. A wakeup was not timed in this round.
One-step velocity lookahead did not remove the wait (gate p50 19.0 ms). Paired near readiness fell 0.79 percentage points; paired far holes rose 0.14 points. Its worst-run p99 was 11.77 ms (limit 11.00) and worst tick 15.11 ms (limit 14.74). After a 90° turn, 252 of 960 retained ahead keys became stale. At 1.4 m/s it compiled 128 ahead tiles per run; paired near arrival was 2.6 ms faster, while far-clear arrival was 14 ms slower. Keep lookahead off.
The queue-only runs also missed the p99 limit (worst 11.46 ms), though their worst tick was 14.19 ms. Neither measured option meets the full bar. Headless 30 m/s completed with zero generation failures (82.71% near readiness, 14.21% far holes); this path has no UE region-completion gate. Geometry checks passed 24/24 for Clang/GCC W1/W8/W12, and the fast tier passed.