Repository navigation
Conversation
|
Independent visual judge: NEUTRAL for all ten startup BEFORE/AFTER pairs and both scoped-world pairs. The amended review confirms complete static UI, terrain, lighting, camera and framing; its earlier animation explanation was withdrawn. Headed NVIDIA Turing / WebGPU. BEFORE: frozen Both complete startup and scoped zero-work assertions pass in browser and the owned desktop host. First-world CPU: 3.0, 3.2, 3.8 ms; median 3.2 ms, against same-harness BEFORE 21.3/22.3 ms. This is render CPU, separately recorded from the first post-READY visible draw. Startup run 1 — frames 1–5
Startup run 2 — frames 1–5
Scoped world — complete resident terrain
Original captures, exact comparison pairs, UI zooms, source patches and runtime receipts. UI zooms are nearest-neighbor enlargements of unchanged source pixels; full original frames remain available. PR stays draft and unmerged. Physical Android/iOS and remote CI are unverified. |
|
Independent review repairs verified: scoped refresh now invokes the world pass once per outer render despite nested bloom quads; repeated scope labels preserve worst first-draw counts and fail closed when any admission has no first draw. Four renderer cases and three scope cases reproduced red before the fixes and pass after them. Opus 5.5/high re-reviewed both and returned FIXED, with no concrete repair regression. The authored world warm-up probe now includes Abyss bloom. Two headed NVIDIA browser runs and desktop pass complete scoped 0/0 ceilings and diagnostics, with all three cells/nine terrain tiles resident. A fresh read-only visual judge returned NEUTRAL (PASS) for each pair and overall against appearance preservation criteria. Immutable repair source patch and runtime receipts. No physical-mobile or remote-CI success is claimed. |


























Startup compiled the raw scene outside the authored output pass, so the first world render rebuilt three warmed graphs and pipelines. Warm-up now exercises the actual output pass before startup and streamed draws, preserving renderer/pass state and the complete offscreen world.
Independent Opus 5.5/high review found two edge cases, now repaired and re-reviewed as FIXED, with no concrete repair regression: forced refresh runs the world pass once per outer render despite nested bloom quads; repeated scope labels retain each metric’s maximum and fail closed if any admission lacks its first draw. Authored update method/property ownership and mode restore in finally. No new API, dependency or Three patch.
Validation:
PRD-616: 100%, 6/6 phases, 5/5 original acceptance criteria, archived under docs/PRDs/done. User authorized pushing repairs and guarded merge queue. Required remote CI must pass; no bypass. Physical Android/iOS launch-time gains remain unverified and external PRD-360 follow-up work. Warm rendering adds bounded work behind the loading cover/offscreen.
Browser CI repair (
70ee3674a): the replay proof now waits for initial startup before stopping/restarting, and its Playwright project uses headed Chromium under the existing CI Xvfb wrapper. Headless SwiftShader lost its device when the newly enabled output warm render attempted to create its swap surface. The unchanged golden replay and 30-second scenario both pass locally (2/2, 1.9 minutes); all 1,800 ticks and every movement assertion are preserved. Root typecheck/lint/docs pass; documentation/CI contracts pass 238/238. Existing lint warnings remain nonfatal. Remote browser CI subsequently passed 4/4.Fresh independent Opus/high screenshot review: all ten real Abyss startup pairs (20 canonical frames, two runs) and both repair terrain pairs are NEUTRAL (PASS) for appearance preservation. The intense green/white terrain glow comes from normal-material terrain plus the deliberately bloom-enabled streamed-world probe; it is the same in both arms. That fixture alone cannot prove real Abyss parity. Real Abyss's lamp, pearls, additive plankton and bloom are intentional; Opus found no changed glow, UI, particles, lighting or framing in its actual startup images. These startup frames cover the attract screen, not later combat. The immutable image links above remain the reviewed evidence.
Remote CI repair (
cec31f309): unit shard 3/4 reached the build-exit test's 60-second ceiling; its log alone cannot distinguish resource contention from a leak. The fixture now retains onlypickup.wav, avoiding unrelated model/texture encoding while exercising the real audio worker pool. The 60-second ceiling and clean-exit assertion are unchanged; a new assertion requires the audio pass to report one processed input. Focused local runtime fell from 28.56 seconds to 880 ms (1.88 seconds including Vitest startup). A real negative control disabling worker-pool disposal still timed out; the ignored compiled file was restored byte for byte.pnpm ci:fastpasses agents/docs/lint/drift on the final repair. Fresh remote qualification remains pending; current-attempt receipt rules require a complete candidate run, so a failed-job-only retry is inappropriate.