Fable (integrator): ROUND 38 CLOSED — v8 wave 1, the town is inhabited; 292/300 stands (the worst frame is night, both new layers are zero at night)

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

docs/LANES/ROUND38_INSTRUCTIONS.md

@@ -112,6 +112,67 @@ dirty; commit around it by pathspec.
    blocks on E).**
 2. **Then F:** the four gates, full `qa.sh --strict --matrix`, round close. No tag.
 
+## ROUND CLOSED (Fable) — **THE TOWN IS INHABITED. All four items, all green, no tag.**
+
+`qa.sh --strict --matrix`: **7 passed · 0 failed · 0 warn · 0 skipped** · flags_check **0/0** with
+all four new gates inside it · **MATRIX GREEN 10 towns × 7 gates**, 0 fetches under `?noassets`.
+
+**THE BUDGET: 292/300 STANDS — margin is still EIGHT, not six.** My naive sum (292 + 2 = 294) was
+wrong, and wrong in our favour: **the pinned worst frame is a NIGHT frame, and both of wave 1's
+draw-spending layers are structurally zero at night** (the washing comes in, the magpie roosts).
+Wave 1's one draw is spent in daylight, where the walked peak sits ten draws lower. F reproduced
+the R37 pin **to the byte, twice** (292 / 120,145 / 31 live chunks) on a control tree, then
+isolated every delta by flag revert: `?yards=0` at night returns all three control numbers;
+`?yards=0` at noon isolates the noon +1 to the washing alone.
+
+**B's control-346 is adjudicated, not merely cited.** As a *delta* it was the wall-clock ped-spawn
+trap and B was right to hide the sim — but ">300" is a claim about the world, and neither existing
+method could adjudicate it (one photographs a half-dressed street, the other arrives clean at 27
+chunks). F built `budget_walk.py --method stepsettle` — fixed-yaw stations, each drained *then held
+N real rAF frames*, so R+1 residue and a full crowd land in one frame. The sign flip is the finding:
+the settle is **+3 draws at noon, −9 at night**, because at night the settled street retires near
+rigs to impostors and a sampler that never waits never sees the crowd go home. **Highest number
+anywhere — 4 methods × 2 trees × 2 segments × 5 towns — is 292. Nothing read over 300.**
+
+**F corrected its own R37 pin:** corpus tri worst is **130,727** (fitzroy), not 124,793 — and it
+reads 130,727 on the *control*, so it is R37's 8 m station spacing vs this round's 2 m, not a wave-1
+cost. *Station spacing is a bigger lever on a walked worst frame than settle policy.*
+
+**What shipped:** 21 distinct palettes across 23 towns at +0/+0 (classic byte-identical **by
+construction** — `townCharacter(null)` returns the v1 literals with the same material cache key) ·
+36 `yard` lots that had rendered as blank shopfront sheds since v1 are now fenced backyards, and
+the box props are **−2 draws** because those lots used to fall through to `buildShopfront` · the
+magpie at +1 draw / 182 tris, season now **23 distinct positions across 24 towns** · six audio
+layers at zero draws **that pan** (B refused the no-panning limitation and added a
+`StereoPannerNode`; F proved the pan is a bearing: yaw 0/90/180/270 → +0.00 / +0.85 / −0.00 / −0.85).
+
+**Three self-caught errors, one per lane — the round's real signature.** B's own season mechanism
+was **vacuous**: keyed on `citySeed` alone it made **24 of 24 towns a constant 214 days** from
+swooping season, and it would have gated green because `?magpie=1` forces the state. F's own pan
+arm used two bearings, read ±0.00, and **would have called a working panner inconclusive** — four
+bearings now. E generated a perfectly good 1,121-tri `paling_fence` that B rejected on measurement
+(905 instances = **1.01 M triangles** against ~75k spare), producing the sizing law this round paid
+for: **an asset's budget is `tris × the instance count the placement rule implies`** — get the
+instance count from the consuming lane *before* generating anything that tiles or repeats per-lot.
+
+**Assets:** `trellis2_mlx` swap verified **1.92×** (and it hid a trap — `sorted(rglob)[0]` was
+picking the *untextured* candidate); six sounds; four grounds; three GLBs published and
+**sha1-verified in both directions off the public depot**; and two **decisive failures** —
+`hills_hoist` and `tv_aerial` cannot be generated (TRELLIS returns thin wire as thousands of
+fragments; no voxel size keeps a 30 mm pole and a 2 mm wire at 600 tris), deleted so nothing could
+catalogue them, permanently primitive at +0 draws. The magpie shipped reading as a **crow**; E
+tinted it by projecting anatomical windows through the UVs in the mesh's own frame (a 100-island
+smart unwrap has no island that announces itself as "nape"), and **Fable verified the render by
+eye**: pied and unmistakable at 64/48/32 px.
+
+**Carried:** two magpies now exist (B's 182-tri procedural bird ships; E's 894-tri tinted GLB is a
+named upgrade at `birdGeometry()`/`birdMaterial()`) — **filed for a side-by-side screenshot, not
+forced** · `magpie.js` carries two stale tri comments (154 vs measured 182) · §1.4 is synthetic-only
+until the civic fetch, and its gate **skips loudly** on real towns rather than passing vacuously ·
+audible output is unverifiable headlessly and every run prints so — *nobody in this loop has ears* ·
+A's D1–D4 · the matrix budget column is bookmark numbers · `?r=3` breaches at 307 · the
+`?classic=1` selector question.
+
 ## Standing
 
 Nothing needs John. Carried debt (R37's ledger, unchanged): A's D1–D4 · the matrix's budget