Lane C measured the bookshelf's transmission halving interior draws (opshop/hall 191->104, book/hall
101->51, dept/hall 116->116 control). The cause, with vendored line numbers: three.module.js:16209
enters renderTransmissionPass(opaqueObjects, ...) whenever currentRenderList.transmissive is
non-empty, and :16433 re-renders the WHOLE OPAQUE LIST into a render target — every opaque draw
issued twice. Trigger is material.transmission > 0 (:6890), set from the glTF at GLTFLoader:1202.
And at 1.0 it is not 'a bit glassy': :494 does totalDiffuse = mix(totalDiffuse, transmitted.rgb,
material.transmission) — the diffuse term is REPLACED by the backdrop.
I SCANNED ALL 103 GLBs. Three carried it:
procity_fit_bookshelf_01 mtl_10218_Bookshelves_v1 shipped and hurting (C's)
procity_street_longbench_01 lambert3SG NOT WIRED — the 5.86 m arcade bench
the v9 charter names, on my R38
recommend list for Lane B
procity_street_streetlight_01 Streetlight_HighResSG3 NOT WIRED — one per ~18 m of street
The two unwired ones are the worse finding: the street budget is 292/300 with EIGHT draws of margin,
and a pre-pass there costs another ~292, not eight. Armed, not fired.
FIXED SURGICALLY. pipeline/strip_transmission.py rewrites only the JSON chunk and copies BIN through,
printing sha1(BIN) before and after: BIN IDENTICAL on all three, tris/dims/mtl/img/tex unchanged.
KHR_materials_ior deliberately left (ior does not trigger the pre-pass).
PUBLISHED + VERIFIED under John's R22 standing authorization via the passwordless tailnet ingress:
6/6 HTTP 200 off the PUBLIC depot, byte counts exact, sha1 identical, depot copies re-scanned clean.
_published.json 56 -> 59.
GATED, which matters more than the fix: validate_manifest.py now HARD FAILS on transmissionFactor>0
with the three.js line numbers and the fix command in the message; glb_stat.py grew a transmission
column and a loud footer. Nobody looked for three epochs because a bookshelf is obviously not glass.
ALSO: the arcade's three meshes, published and catalogued the same round they were generated —
aframe_board 366 tris (<=16 inst), blade_sign 299 (17 inst), keycutter_sign 498 (1 inst). manifest
furniture 13->16, GLBs 36->39, validate 0 errors 0 warnings. blade_sign ships FLAGGED: its back face
carries TRELLIS-hallucinated art where the concept had no data, which is one more reason the
+0-draw perpendicular-quad answer is the recommendation and this GLB is the fallback.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A. DEPOT — closed, under John's R22 standing Lane-E authorization (the authority that
shipped sit.glb in R28 and the nine R36 sweep assets). Pushed via the passwordless
tailnet ingress: _published.json 53 -> 56, drift check clean. Then verified the way
this lane verifies — FETCHED BACK OFF THE PUBLIC DEPOT and hashed, not "the upload
said 200": magpie/paling_fence/water_tank all HTTP 200, byte counts exact, sha1
identical to local, 3/3. _r38_results.PENDING_PUBLISH.json renamed, build_manifest
re-run (GLBs 33 -> 36), validate_manifest 0 errors / 0 warnings.
B. THE MAGPIE TINT — done, not offered. The bird shipped as an all-black crow (R38 4d
traced the loss to trellis2_mlx's PBR bake). "Hand-tint the atlas" is not a paint job:
the bake is a smart-UV unwrap of ~100 islands and none of them announces itself as a
nape. So pipeline/tint_atlas.py defines the tint in the MESH's frame and projects it
through the UVs — rasterize every UV triangle to texel->(position,normal), evaluate
soft anatomical windows on the positions (white nape behind a black crown, shoulder
bar on the flanks only so the dorsal midline stays black, rump + tail base stopping
short of the black terminal band), paint keeping the bake's own luminance as shading,
then dilate 5 texels into the UV padding so no black fringe survives bilinear/mip.
3-D masks give one consistent soft edge across every island at once; a 2-D blur could
not (it would bleed island-to-island through the black background).
GEOMETRY UNTOUCHED byte for byte: still 894 tris / 1 mtl / 1 img / 512^2 WebP / no
Draco / 0.155x0.358x0.404 m. Painted luma median 0.84 (p90 0.95) vs the bake's 0.13.
VERIFIED IN THE GAME'S OWN RENDERER: the PUBLISHED file loaded off digalot.fyi through
the vendored GLTFLoader + three.js — 1 mesh, 894 tris, 512x512 texture decoded (which
also proves the PIL-written WebP decodes in a browser).
VERDICT: it reads. Pied and unmistakable at 64 and 48 px, still black-and-white at 32.
Honest caveat: from directly below-and-in-front it stays dark — correct, a real
magpie's bib and belly ARE black. Render: docs/shots/laneE/r38_magpie_tint.png.
Standing after B's 4514cb9: an UPGRADE PATH, not a dependency — B's 182-tri procedural
bird is 4.9x lighter and the right call. Nothing probes for the GLB and nothing should.
C. CORRECTIONS FOLDED INTO THE RECORD so the next asset round does not re-learn them:
skins are 9.7 s not 5.7 s (1024^2, 1.78x the pixels, per-pixel identical) - the reuse
list's "one line each" holds for 5 of 9 (extractGLB keeps only the FIRST material) -
normalize.py's collapse decimator floors at the shell count on TRELLIS output while
bake_lowpoly.py clears it every time - the 489 s outlier at 3.8x median means
throughput plans on a tail, not a mean - furniture.js:30's stale comment is fixed by B.
NEW, and the one worth keeping: an asset's tri budget is tris x THE INSTANCE COUNT THE
PLACEMENT RULE IMPLIES. paling_fence is a good 1,121-tri panel and a rejected asset —
905 instances = 1.01M tris, ~47x over. Ask the consuming lane for the instance count
BEFORE generating anything that tiles, edges or repeats per-lot.
New tools, all on-device and $0: tint_atlas.py, render_views.py (view sheet + --eevee +
roster-consistent --thumb), view_sheet.py (the distance strip no 256px thumbnail can make).
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A. PIPELINE FIX: gen_props.py trellis_mac -> trellis2_mlx, contract read from its
manifest.json. The two operators are NOT drop-in on output — trellis2_mlx writes
BOTH candidate.glb (raw) and candidate_pbr.glb (baked), so the old
sorted(rglob('*.glb'))[0] would have silently shipped untextured meshes. Added
_pick_glb() (result.json first) + the bg_remove_local cutout stage, fail-soft.
MEASURED: median 128.0s vs trellis_mac's 245.4s p50 = 1.92x. 5/5 first attempt,
0 retries against the 4-5% budget. One 489s tail outlier, no error — named.
B. AUDIO (1.1) — six PACK entries, +0.86MB, ZERO draws/tris, all originalby
construction: dog-bark-fence, mower-distant, radio-through-window, sprinkler,
roller-door, magpie-carol. Seeded via crc32(name); beds fold seamlessly
(wrap-jump <=0.18x internal step). Verified by spectrogram + band-energy, and
one defect found by LOOKING: the sprinkler's spray masked its own ticks and read
as a hose — rebalanced and re-rendered. Distance is spent in the SPECTRUM, not
the level, because encode()'s loudnorm undoes a quiet render.
C. GROUNDS (1.2) — gen_skins.py had NO ground template; added one with its own
outdoor style prefix (the house STYLE says "warm fluorescent", an interior cue).
4 new 512^2 square tiles. Review's 5.7s is 9.7s here — same per-pixel, mine are
1.78x the pixels. Found: gravel/reddust are selected by nothing partly because
their `use` values name no slot ground.js:177 has.
D. MESHES — 3 shipped (magpie 894 tris, paling_fence 1121, water_tank 500), all
GLB-law clean. 2 FAILED and were DELETED: hills_hoist + tv_aerial. Thin wire
returns as thousands of disconnected fragments, so collapse floors at 250k/89k
AND a remesh coarse enough for 600 tris erases the pole. Lane B's boxMats
primitives cost +0 draws and are strictly better. The magpie's geometry passes
but TRELLIS crushed its white nape/wing to black — traced to the operator's bake
(the cutout and our decimation are both exonerated); flagged, not hidden.
E. REUSE — all 9 idle assets re-measured; the review's "one line each" holds for
only 5. furniture.js's extractGLB keeps only the FIRST material, so park_bench/
coffee_table/box_crate render in material #1 and bench_wood (0 materials) falls
to grey. Nothing was regenerated that already existed.
The 3 GLBs are staged, NOT published: a depot push is outward-facing and had no
in-session authorization. Held at _r38_results.PENDING_PUBLISH.json; publish +
rename + rebuild lands them. Gate green: validate_manifest 0 errors / 0 warnings.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
B's diagnosis nailed it: 149,161 of 163,015 tris (91.5%) at katoomba venue_night were four of my GLB
street props. Buildings 348, posters 222 (A's cap worked, never the lever), pools 0. B's sharpest point:
it is NOT the density widening -- road density drives the bench/bin/shelter cadence (~100 props on a
966-edge real graph).
Fix (collapse-decimate via the R13 instrument tool; originals kept as _hi):
bench 1942 -> 500 rendered 60,202 -> 15,500
bus_shelter 7998 -> 2500 rendered 55,986 -> 17,500
bin 3000 -> 800 rendered 18,000 -> 4,800
food_cart 4991 -> 1000 rendered 14,973 -> 3,000
furniture total 149,161 -> 40,800 (-108,361)
Projected view ~54.7k (B's 163,015 baseline) / ~113.6k (vs F's 221,935) -- PASSES the 200k gate either
way with 86-145k spare. F: these are projections off B's per-instance attribution; yours is the real
measure. -> F: RE-MEASURE.
Eyeballed every prop; two needed re-cutting: bus_shelter @1000 shredded (thin roof plate on a
voxel-remeshed shell) -> @2500 clean; bin @400 creased -> @800 clean. food_cart's black base slab is in
the SOURCE model (confirmed against the pre-diet thumb from git), not a decimation artifact.
Zero blast radius: dims/footprints/heights identical -> manifest.json byte-identical -> B's placement and
A's goldens untouched. Published 4 diet + 4 _hi archives to the depot (John re-authorized explicitly;
GLBs are git-ignored so the depot IS the archive).
-> Lane B (your 2nd note, cause found, NOT fixed by design): the MeshPhysicalMaterial on bench+food_cart
is KHR_materials_specular in the glTF, and it survives a Blender re-export. Fix = neutralise specular
pre-export. Deliberately deferred to v4.x: the gate is tris (cleared), and changing shipped shading in
a release round without a visual re-verify is risk for no gate benefit. Same list: food_cart is 3.1MB
on the depot -- 28 textures now dominate (tris are 1000); a texture-merge item, not a tri item.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
v2 features are procedural (no forced GLB needs), so E2 is light per Fable's "no
unconsumed entries": two synthetic depot props with clear homes — council-bin ->
procity_street_bin_01 (Lane B street dressing) + milk-crate-stack ->
procity_fit_crate_stack_01 (Lane C floor piles / dig). Both 3k tris, base-origin,
re-normalized to procity names + thumbs, published, verified loading, validate --depot
0 err. Manifest 18 fittings + 10 furniture (28 GLBs). Also noted B3 resolved
(novelty_record 26.5k->8k via E1 bake).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Audit -> normalize (16 GLBs + thumbs) -> web/assets/manifest.json (the
contract Lanes B/C/D read) -> facade gap-fill.
Step 4 facades generated locally on the M3 Ultra via MODELBEAST
flux_local (FLUX.2-klein-4B / MLX on MPS): 23 skins, free, ~5.7s each,
blank signboard, eyeballed, harvested, mapped to shop types. Every
registry type now has >=3 facades (stall 2->4, closing the Lane A gap).
Corner-lot -side walls tagged face:"side" with empty types so no shop
front ever selects a windowless wall. manifest validates green.
Still gated on John (both optional -- the game runs asset-free):
publish the 16 GLBs to 3GOD (needs GOD3_PW), MeshGod hero props (~$2.70).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
