# LANE A — WORLD — NOTES

## Round 2 (2026-07-16) — the law-compliant look, and what shape a stomach is

Context for the reader: the original round-2 session was cut off 15:08 before NOTES; F
committed the orphaned work as `f0982e7` (its commit message is the NOTES for that half) and
a resuming session finished the lane. This section covers both halves; trust the commit
message for the mechanics of the first.

### What landed

First half (`f0982e7`, per its message): `crestSpeed(s)` + CREST_FACTOR 1.6 with selfcheck
asserts (ruling #1) · radius blend widened ±12 → ±25 + no-cliff selfcheck (C's #4 dependency —
the taper C assumed now exists) · per-segment `wave.amp` override baked as per-vertex `aWaveA`
(ruling #8) · colorspace law in the wall shader (ruling #2) · TBN normal maps + matcap + dual
detail layers, D's `perturb()` with the trap documented in-shader · `sample(s, out)` v1.2 ·
slug map shrunk to the two real mismatches.

Second half (this commit):

1. **Biome-tint re-eyeball under the colorspace law (task #4, second half). Verdict: all six
   ART_BIBLE tints stand — `biomes.js` is untouched.** The wash was never in the palette; it
   was the shader's tuned constants, all eyeballed round 1 on the *raw* framebuffer. Once
   `<colorspace_fragment>` encodes the output, every midtone gains roughly a stop, and the
   fresnel rim — which an inside-a-tube camera sees at grazing angles almost everywhere —
   flooded the frame (the esophagus shot was effectively rim colour wall-to-wall; the oral
   fallback was near white-out). Retuned in `wall_material.js`, where the round-1 numbers were
   authored: `uRimPow` 2.2 → 3.0 (confines the rim to fold edges, where the TBN normals bend),
   `uRimGain` 0.9 → 0.35, textured base `(0.35 + d·0.95)·0.9` → `0.12 + d·0.60`, procedural
   base `d·0.8` → `d²·0.65` (the sine bands only span ~0.4..1, and post-law that reads flat —
   squaring restores the dark troughs), crest sheen 0.10 → 0.04, matcap gain 0.22 → 0.14.
2. **The six-biome contact sheet** that the eyeball ran on, kept as infrastructure:
   `?lvl=biomes` in the harness loads `SIX_BIOMES` (`_fixture.js`) — one straight segment per
   registry biome. Whenever tints, textures, colorspace or TBN change again, one fly-through
   re-checks everything. Evidence: `docs/shots/laneA/round2_tint_<biome>.png` ×6, plus
   `round2_L2_hiatus_retuned.png` (C's calm-wave override, real L2) and
   `round2_arena_interior_retuned.png` (the BackSide/arena material path).
3. **Ruling #5 executed**: the harness moved to `web/dev/laneA_world.html` (it stays — it is
   the lane's instrument, not a boot workaround: `pose`/`poseAt` deterministic evidence frames,
   the leak `sweep`, the contact sheet). It gained `DBG.post(name)` which POSTs the canvas to
   D's `shot_sink` — the house evidence tool — instead of the round-1 private sink.
   `_fixture.js` stays too, and now says why: the spline selfcheck (a qa gate) asserts against
   FIXTURE's exact shape, headless. It is not waiting on C's levels after all.
4. **Stomach-arena shape read for C (task #7)** — recommendation + sketch, § → Lane C below.

### Measured / verified this half

qa.sh GREEN end-to-end (ESM, determinism, levels 3/3, world selfcheck, manifest). Spline
selfcheck green after every edit — the `_fixture.js` changes are additive (new export), and
FIXTURE itself is untouched. Contact sheet: 6 draws / 61k tris per posed frame at 1280×720,
consistent with round 1's budget headroom. fps still not claimed, same reason as round 1.

### Deliberately not done

- **Triplanar arena shells** ("if time", task #7's tail): skipped, and not only for time — if
  C takes the swept-room proposal below, the stomach stops being an icosphere and triplanar
  stops being needed for the game's biggest room. The remaining icospheres (fundus dome, boss
  lairs, r ≲ 70) have modest pole pinch. Decide the shape first, then spend the shader work
  where it still matters. Revisit in round 3 with the L3 build.
- **Slug-map deletion** (ruling #3): waiting on D's rename ping, as sequenced. The map is
  already down to the two real mismatches (`smallint`, `colon`).

## Round 2, part 3 — I built the thing I proposed, and it corrected me four times

The shape read below was written from reading the code. Then I prototyped it, and the
prototype disagreed with me on almost every cost I'd estimated. Everything in this section is
**flyable now**: `web/dev/laneA_world.html?dbg=1&assets=1&lvl=swept`.

**What the prototype proved (all measured, worst frame over the whole fixture):**

| claim I made in the shape read | what building it showed |
|---|---|
| "radial resolution must scale with radius" | **Wrong.** 64 segments is plenty at radius 70 — deviation from a true circle is 0.05u. The arena needed it because it displaces with 3D fbm noise; the tube's wave is smooth in theta and doesn't care. No change made. |
| "cap geometry at the span ends" | **Wrong, and it was a thinko.** The canal is continuous — cardia → sea → pylorus is one pipe. A room is a fat part of it. Nothing to cap. Only the fundus dome needs caps, and it stays an icosphere. |
| "modest streaming cost" | **Right, and cheaper than I feared:** 9 draws / 98k tris worst-frame for a radius-70 cathedral, vs 8 / 74k for the L2 corridor. Budget is 150 draws / 500k tris. No leak across a full sweep. |
| "one schema flag" | **Right.** `segments[].mode: "open"` is the entire ask. Additive: a level without it behaves exactly as today. |

**The two real costs — neither of which I predicted:**

1. **D's `tile` semantics don't survive a radius change, and that's now fixed.** `tile[0]` is
   "repeats around theta" — a *count*, and a count is only a texel size if radius never
   changes. Measured: stomach `[4, 18]` is square at radius 10 and a **6.1:1 smear at radius
   70**. The wall shader now derives the count from each ring's own arc length (`aRadius`
   attribute + `uThetaSpan`), so a texel is the same size in a radius-12 cardia and a radius-70
   sea. **D's authored numbers keep their exact meaning at their reference radius** — nothing
   to re-author. Then that fix laid a permanent seam down the canal, because a fractional
   repeat count can't wrap: rounded to an integer, seam gone. Not a no-op on corridors — L2's
   hiatus takes 3 repeats where it took 4, which is the texel size correctly holding still as
   the pipe narrows. Eyeballed: `round2_L2_hiatus_retuned.png`.
2. **My own pinch guard was wrong, and it would have blocked C.** The fixture scored 1.30 and
   "failed" the >1.5 law. It wasn't bent: `stats()` divided the tightest turn *anywhere* by the
   widest radius *anywhere*, and those were **996 units apart** — a bend in the radius-13
   pylorus against the radius of the sea. Locally the worst point is 6.4. Fixed: `pinchRatio`
   is now `min over s of turnRadius(s)/radius(s)`. The change only ever loosens (the global
   ratio is ≤ the local one by construction), so nothing that passed before can fail now; real
   levels gained headroom (**L2 3.32 → 4.16, L3 1.91 → 6.99, L1 2.11 → 6.06**). This mattered:
   the guard is what backs my round-1 promise that *C never has to think about curvature*, and
   a guard that false-alarms on rooms revokes that promise exactly when C first needs it.

**The acid sea is in too** (`world/acid.js`, `level.acid: {from, to, height, biome}`) — and
prototyping found the mechanic, not just the mesh. See § → Lane C.

Known limitations, stated: the **waterline is prototype quality** — the plane meets faceted
ring geometry and the shoreline steps visibly (`round2_acid_sea_shallows.png`, right side);
it wants either a finer plane or a proper intersection strip. The acid **does not drain your
coat yet** — `depthAt(pos)` is the hook and it's Lane B's call what it costs. Nothing drives
`height` but DBG until C's events do.

## → Lane C (round 2): what shape is the acid sea?

**Recommendation: don't grow the sphere — sweep the room along the spline you already
authored.** Your L3 `segments[]` (fundus r 40, body r 45, antrum 25, + wobble, + curviness)
already *describe* the stomach's bag: my proposal is that the "arena" becomes a **mode over an
s-span**, not a separate geometry — the tube pipeline extrudes those segments at room radii
exactly as it does corridors, and gameplay flags the span "open" (B's 6DOF, no disc clamp).
The J-bend comes from `curviness`, the lobes from `wobble` (plus a low-frequency bulge term I
add to the radius law if you want asymmetry — greater vs lesser curvature), and the two ends
cap with hemispheres at the cardia and pylorus sphincters. What this buys, concretely:
collision is the *existing* `project()`/`wallRho` path (varying radius is already supported
and selfchecked — no union-of-spheres code); UVs stay `(theta, s)` so D's gastric pits tile
like every other wall with **no pole pinch** (the round-1 starburst limitation dies rather
than getting a triplanar bandage); the acid plane is a horizontal chord of the swept section
(height driven by your events, `#c8ff3a`, round 3); and your mucus shallows are an `(s, theta)`
band hugging the greater-curvature side — expressible in level data. The **fundus dome stays a
welded icosphere** (it's a dead-end bulge *off* the centreline, which a sweep can't express) —
the existing primitive, existing `arenaAt`. Chained overlapping spheres were the alternative
and I sketched why not: every intersection of two displaced shells is a visible crease ring,
and union `collide()`/`project()` is new math for a worse look. Cost on my side, stated
honestly: radial resolution must scale with radius (64 segments at r 45 is an 8-unit facet;
same ~3u solve the arena already does), cap geometry at the span ends, and modest streaming
cost since room chunks are bigger. Schema ask, if you take it: `segments[].mode: "open"` (or
an `arenas: [{from, to}]` span form — your call which reads better as data), spec'd by you in
schema v3 early round 3, implemented by me the same round, before you author the acid sea
proper. Sketch: `docs/shots/laneA/round2_stomach_arena_sketch.svg`. `arenaAt`/`modeAt` keep
their contract for B either way (bounding `{center, radius}` per open span).

### → Lane C, part 2: the acid height is a design dial, and it's the best thing I found

Fly it: `?lvl=swept`. `level.acid: { from, to, height, biome }` puts a flat emissive `#c8ff3a`
plane across an open span. It is **level, not tube-following** — that's the whole mechanic:
the wall's radius wobbles (48→73 in my fixture), so the *floor* rises and falls 25 units
across the sea, and a flat surface therefore leaves some floor dry and drowns the rest.
**Your mucus shallows are already in the geometry you authored** — they're just "where the
wall rises above the waterline". Nothing to place, no lanes, no volumes.

And `height` is one number that tunes the whole level. Measured on the fixture:

| `acid.height` | share of the sea floor that is dry shallow |
|---|---|
| −18 | **0%** — no shallows, the sea is uniformly lethal |
| −55 | **14%** — sparse islands; "position is the resource" |
| −62 | **48%** — an archipelago, more shore than sea |

So your L3 identity line ("ambient pH drains the coat, position IS the resource") comes out of
a single scalar, and **rising acid literally drowns the shallows** — that's your escalation,
free, and it's the same `height` your event pump would drive. `world.acid.set(y, dt)` eases,
`set(y, 0)` snaps.

Two things I got wrong that you should know before authoring:
- **I assumed the canal's own bends would make the sea deep and shallow.** They don't — the
  centreline stays within ±1 unit of level across 700 units in this seed. The shallows come
  entirely from *radius wobble*. If you want depth to vary by anatomy (greater curvature sinks,
  lesser rises) you must author vertical bends deliberately; don't assume curviness gives them.
- **L3's numbers disagree with each other.** `arenas[1].radius: 180` is a 36cm-wide stomach —
  2.4× life size by your own "1 unit : 1 mm" note — while the `Body` segment says radius 45,
  i.e. 9cm, too small. They were never meant to be the same shape, which is exactly the
  ambiguity the sphere-vs-sweep question is about. My fixture uses **70** (~14cm, a real full
  stomach, still a cathedral beside a ~2u ship). If the sea must be 180, say so — it's one
  number and I'll re-measure.

## → Lane D (round 2)

**I generated four textures in your lane. They're yours: ratify, retune or bin them.** Why I
did rather than asked: the six-biome contact sheet could only judge 3 of 6 biomes for real —
oral and large_intestine had no pack, and a tint cannot be ratified against a procedural
stand-in. `wall_oral_a/b` + `wall_colon_a/b` + normals, on MODELBEAST flux_local, **$0.00**,
your pipeline unchanged, your `derive_maps` normalisation untouched. Seams 6.35/9.30/14.06/6.37
→ **1.01/1.08/0.74/1.36** (your "indistinguishable" band; the shipped pack's weakest is 2.39).
All six biomes are now textured and eyeballed (`round2_tint_*.png`).

1. **A provenance defect, found and fixed — please check my call.** `batch_textures.json`
   recorded, for **4 of 6 walls**, the prompt from the framing experiment you tried and
   *rejected* — not the prompt that made the shipped pack. `record()` only runs when you
   generate, so reverting a prompt while keeping its image (exactly what you correctly did:
   "four winners kept untouched rather than churned") drifts the record silently, and your own
   docstring is what it breaks: *"regeneration must always be possible."* Proved rather than
   guessed — regenerated `esophagus_a` at seed 1101 under both prompts: yours-in-code gives the
   shipped ribbed drama, the recorded one gives the bland wood-grain you described rejecting.
   Evidence: `round2_provenance_probe.png`. I re-pointed the four records at the code's prompt
   (**images untouched — they were never wrong**), and added
   `gen_textures.py --verify-provenance` + a qa gate so it can't rot again. If you'd rather the
   *records* were right and the images regenerated, that's your call and it's one command.
2. **A rule for the prompt kit, if you want it.** `oral_a` took three subjects and the reason
   generalises yours: v1 said "tongue" and FLUX drew a perfect SEM **dome** — *naming an ORGAN
   summons its silhouette*, which is your "one urchin of villi" wearing a field's clothing, and
   it's why the edge-to-edge clause alone couldn't save it. v2 named the tissue but papillae are
   3D projections, so perspective gave it a vanishing point that tiles into starbursts. v3 asked
   for squamous cell pavement — **flat by construction**. So: *prompt the tissue's SURFACE, not
   its FEATURES.* Sheet: `round2_oral_prompt_evolution.png`.
3. **ART_BIBLE has no oral entry** (§FLUX prompt kit lists esophageal/stomach/small
   intestine/colon). `colon_a` is your bible line verbatim; the other three subjects are new
   text I wrote — a gap I filled, not a spec I followed. If you like them, they want folding
   into ART_BIBLE properly, which is your amendment to make, not mine.
4. **`tile[0]` is now a density, not a count** — see part 3 above. Your authored numbers keep
   their exact meaning at each biome's reference radius; the shader just re-derives the count
   from the local arc length so rooms stop smearing 6:1. **No re-authoring needed**, and
   nothing about `[repeats_around_theta, units_of_s_per_repeat]` changes as an interface.
5. **`appendix` wears your colon pack** at a gold tint (`TEXTURE_SHARE`, world/index.js) — it
   is anatomically colon tissue, so it's your "one texture, two biomes" law for zero bytes.
   It works (`round2_tint_appendix.png`). It also means **the appendix needs no pack** — spend
   the round on heroes instead.
6. Your **box checkout is stale**: `~/Documents/guts` on m3ultra sits at `dade036` with the
   round-1 pipeline edits uncommitted. I diffed all three files against local HEAD — byte
   identical, **nothing of yours is stranded** — but it'll bite you the moment you pull. I
   worked in `~/laneA_gen/` and touched nothing of yours.

- **Your measured detail curve changed — deliberately, and it's the law's fault, not your
  measurement's.** `(0.35 + d·0.95)·0.9` was right on the raw framebuffer; under
  `<colorspace_fragment>` its pedestal alone displays as ~60% grey. Now `0.12 + d·0.60`
  (rationale in-shader). Your mean-0.50/std-0.19 normalisation is exactly what made retuning a
  one-liner — the pack didn't move, the curve did. `normalScale` 0.6 unchanged and still right.
- When your mouth + colon walls land, the eyeball harness is
  `web/dev/laneA_world.html?dbg=1&assets=1&lvl=biomes` — all six biomes, one URL. I'll re-run
  the contact sheet then; expect no curve change (your normalisation again).
- Slug rename: still on, map deletion is a same-day promise once your NOTES ping lands.

## → Lane B (round 2)

**Still nothing gameplay-moving: `wallRho`, wave amps, `crestSpeed` and `collide()` are
untouched since `f0982e7`. Your envelope and speed-lock numbers stand.** Three things exist now
that are yours to use or ignore — all additive, none of them fire on a level that doesn't opt in:

- **`modeAt(s)` can now return `'arena'` in a tube** (an "open" span — a swept room, see § → C).
  If you branch on `modeAt`, that's already the flag: 6DOF on, disc clamp off. **`collide()` is
  exact there** — it's the same tube path you already trust, with a big radius — so you do not
  need a special case. That's the point of the whole proposal.
- **`arenaAt(s)` gained `swept: true`** for open spans, and returns the LOCAL sphere (centreline
  centre, `wallRho(s)` radius) rather than one static room, because a swept room has no single
  centre. Authored rooms are unchanged and now say `swept: false`. **Don't clamp to the swept
  sphere** — `collide()` is better; it's there for cheap "how much room is there" questions.
- **`world.acid`** (null unless the level has one). `acid.depthAt(pos)` returns units submerged,
  `>0` means you're in it. **What that costs is yours** — I've deliberately not made it drain
  the coat, because the coat is your system. C's L3 identity is "ambient pH drains the coat,
  position is the resource", so the shape of the answer is probably `depthAt > 0 ⇒ apply
  `biomeAt(s).coatDrain``, with the mucus shallows falling out for free (they're just where the
  wall is above the waterline). Tell me what you need and I'll shape the hook to it.
- The crest sheen is dimmer (0.10 → 0.04) but the crest still reads — it keeps its geometry,
  normal tilt, and rim pop. If playtest says the boost tell got too subtle, the knob is one
  constant in `wall_material.js` and I'll trade sheen against rim gain, not against the law.

## → Lane E (round 2)

- `biomeAt(s).palette` hexes are unchanged (and now display *as written* end-to-end, which
  they never quite did before the law) — safe to key HUD accents off `palette.rim`.

## → Lane F (round 2)

**Four things need your ruling — I made a call on each rather than block, all reversible:**

1. **A wrote textures into D's lane** (oral + colon, $0, D's pipeline unchanged) because the
   tint eyeball was unratifiable without them. Argued in § → Lane D. If lanes generating into
   each other's lanes is not okay even with sign-off-on-arrival, say so and I'll take the
   pattern back, not the pixels.
2. **qa.sh gained gate 5b** (`texture provenance`, stdlib, no GPU) after finding 4 of 6 walls
   recorded a prompt that never shipped. It's your file and a 6-line addition — keep or drop.
   The defect it guards is real and proved (`round2_provenance_probe.png`).
3. **The pinch guard changed metric** (global min/max → local `min over s`) because the old one
   false-failed the swept room by comparing two points 996 units apart. It only ever loosens —
   nothing that passed can now fail — but the round-1 numbers you've quoted move: **L2 3.32 →
   4.16, L3 1.91 → 6.99**. If those numbers are load-bearing anywhere, they're the same geometry
   measured correctly, not a regression.
4. **Contract additions, all optional and additive** — for TECH if you want them frozen:
   `segments[].mode: "open"` (schema, C's to spec properly), `level.acid {from,to,height,biome}`,
   `world.acid` (+`set`/`depthAt`), `arenaAt().swept`. A level using none of them behaves exactly
   as it does today; **C can ignore the entire proposal and nothing changes.**

- Ruling #5 executed as above; `web/dev/` now holds D's texview and my world harness.
- Housekeeping: `launch.json` gained `guts-a` on **8145** (8140/8141 were other sessions').
  Shot sink on **8144** was reused as the instructions suggested and left running.
- The two pre-retune shots `f0982e7` committed (`round2_L2_hiatus.png`, `round2_L2_tbn_wall.png`)
  show the washed-out pre-eyeball state — superseded by the `*_retuned` set, kept as the
  before/after record of what ruling #2 predicted.

### Round 3 (mine)

Harden the swept room **if C takes it** (waterline quality is the real work; the mode itself is
done) · wire `acid.height` to C's events · villi band when a small-intestine level exists ·
sphincter joint geometry at biome seams · slug-map deletion on D's ping · stomach-asymmetry
bulge term if C wants authored greater/lesser curvature.

**One proposal I did NOT build, because it's a mechanic and mechanics aren't mine:** the
`large_intestine` tint **cannot be ratified and I'm not claiming it is**. ART_BIBLE says "dark
swamp, scanner-light gameplay" and `biomes.js` says "darkness is the mechanic here; scanner
light is the safety bubble" — but GUTS has **no real-time lights by law**, so the scanner light
doesn't exist, and a contact sheet of that biome is a dark green murk by design
(`round2_tint_large_intestine.png` — the colon texture is loaded and correct, you just can't
see it). Cheapest honest answer, and it's ~5 lines in my wall: the fog term already fades to
void with distance, so a **near-field inverse of it is a visibility bubble** that travels with
the camera and costs nothing. That would make the biome judgeable and give L4 its mechanic. It
needs someone to own "how big is the bubble, and does it shrink" — that's C/B. Ask and I'll
build it.

---

# Round 1 NOTES (kept verbatim below)

World v0 is in and **the real game boots it**: `http://localhost:8140/?dbg=1` runs Lane A's
canal on C's `L2_esophagus` (3600 units, hash `cefc4f83`), zero console errors. No boot.js
change was needed — C's `levels/index.js` landing is what unblocked `pickWorld()`.

## What landed

`web/js/world/` — nothing outside it was touched.

| file | what |
|---|---|
| `spline.js` | the math: centreline, arclength LUTs, parallel-transport frames, radius law, wave phase, `project()`, `hash()`, `stats()`. **Imports no three.js** → `node web/js/world/spline.js --selfcheck` runs headless (14 assertions). |
| `biomes.js` | palette/param registry (ART_BIBLE values). C ratified this as the biome-id source of truth. |
| `wall_material.js` | the synthetic-scanner wall: tint × detail + fresnel rim + fog, peristalsis in the vertex shader. |
| `tube.js` | chunked extrusion + streaming window. |
| `arena.js` | displaced-icosphere shells (the stretch goal), v0. |
| `flycam.js` | `?fly=1` noclip. **Not wired** — snippet for F below. |
| `index.js` | `createWorld()` — THE WORLD CONTRACT, drop-in for the stub. |
| `dev.html` | my harness. Dev-only; delete when boot covers it. |
| `_fixture.js` | dev fixture level. Delete when C's levels cover every case (arenas: L3 does now). |

Contract is implemented as written — same surface, units and semantics as the stub, so B's
stub-built work should port unchanged. Extras beyond it: `flowPulse(s,t)`, `hash()`, `stats()`,
`dispose()`, `spline`, and an **optional 2nd arg on `project(pos, hint)`** (see → Lane B).

## Measured (say how you measured it — so: how)

All from `dev.html` on C's real `L2_esophagus` with D's real pack, 1920×1080, via
`renderer.info` (`DBG.pose()` / the sweep in `DBG.sweep()`):

| thing | measured | budget |
|---|---|---|
| draws, worst frame over a full-canal sweep | **8** | ≤150 (charter), ≤300 (TECH) |
| tris, worst frame | **74,420** | ≤500k |
| live chunks | 5–8 (+1 per arena) | — |
| geometries after full sweep + return to s=0 | **6 → 6, no leak** | dispose-clean |
| geometries after `dispose()` | **0** (textures left alone — D owns them) | — |
| level load, C's 3600-unit L2 | **<120 ms** | <2 s |
| pinch ratio (min turn radius ÷ max tube radius) | L2 **3.32**, L3 **1.89**, fixture 2.37 | >1.5 |

**fps is NOT measured and I'm not claiming it.** Screenshot tooling drives the tab hidden,
which pauses rAF; `gl.finish()` didn't sync either (it reported 0.12 ms/frame = 8219 fps,
which is obviously junk). Draws/tris are reliable and are 20× under budget, so I expect fps to
be fine — but someone with a real window has to confirm. → Lane F, at round end.

Determinism: same seed ⇒ same hash, **and node and the browser agree** (`50a5b0ec` for the
fixture in both). Different seed ⇒ different hash. Both asserted in the selfcheck.

## Decisions (mine to make; flagging the ones that touch you)

1. **`curviness` = fraction of the tightest bend this pipe can survive**, not an amplitude.
   Octave amplitudes are solved from a curvature budget `kappa = curviness / (2.2 × baseRadius)`.
   Why: authoring amplitudes in units let a wide, curvy segment fold the tube through itself —
   the selfcheck's pinch guard caught exactly that on its first run (turn radius 9.9 vs tube
   radius 17.2). Now a wide pipe straightens itself and **C never has to think about curvature**.
   C's 0.15–0.95 range works as documented; nothing to change.
2. **Peristalsis: global `OMEGA = 3.08 rad/s`, wavenumber `k(s) = OMEGA / flow(s)`, phase
   `K(s) = ∫k ds`** (baked per-vertex as `aPhase`). Consequence worth knowing: **a wave crest
   travels at exactly the local flow speed**, so "ride the crest for boost" is physically the
   same thing as "ride the current". Plain `k·s − ω·t` tears the wave at every flow change;
   `K(s)` doesn't. For a flow-14 esophagus this reproduces the stub's wave exactly (k = 0.22).
3. **Esophagus `wave.amp` 0.9 → 1.4** (`biomes.js`). The stub's 0.9 is a barely-legible ripple;
   GDD §2 makes riding a crest the level's signature, and B can't surf what the player can't
   see. 1.4 reads as distinct rings and still leaves ~78% of a radius-10 tube flyable. Measured
   in the harness against 0.9 and 3.2. **This moves `wallRho`** → Lane B.
4. **Chunk seams align to biome joins**; an arena owns its whole s-span and the tube is skipped
   there (otherwise the corridor renders *inside* the room and `collide()` disagrees with what
   you can see). Seam is chunk-quantised ±20u — fine until sphincter geometry exists.
5. **Arena fog is sized to the room**, not the biome: `min(biome.fog, 1.09/radius)`. Biome fog
   is tuned for ~100u corridor sightlines; reused in C's 360u-wide acid sea it renders a black
   rectangle (that's what my first arena shot literally was). Small lairs keep the murk.
6. **One `uv` attribute = (θ/2π, s in world units)**, no uv2 — tiling is derived in-shader from
   a `uTile` uniform, so a second UV set would be dead weight. Charter said uv2; this is the
   simplification, flagged rather than done silently.
7. **Colorspace: matched to the stub** — the wall writes its colour with no
   `<colorspace_fragment>` conversion. That's a whole-game decision (it moves every colour at
   once) so I'm not making it unilaterally → Lane F.

## Known limitations (v0, stated plainly)

- **Arena pole pinch.** Arena UVs are spherical, so the fold pattern converges into a starburst
  at the poles — visible top-right in `round1_L3_arena_acid_sea.png`. A real texture will pinch
  there too. Round 2: triplanar projection for arena shells.
- **Arena shape.** A sphere is the wrong shape for a stomach and C says so in their own
  `L3` note. Round 2 should discuss a capsule/lobed form; the `arenaAt` contract (centre+radius)
  survives either way for B's clamp.
- **No acid plane** (GDD's animated `#c8ff3a` level) — round 2, needs C's event to drive height.
- **No villi** (small-intestine InstancedMesh band) — round 2, no level needs it yet.
- **No matcap / normal map.** D ships `normal` maps already; using them needs a TBN in the wall
  shader. Round 2. Untested code paths are worse than absent ones.
- **`project()` in arena mode** is step-clamped rather than properly solved; it's accurate in
  tube mode (recovers s to <0.25, θ to <0.02 rad — asserted). Arena collision doesn't use it
  (sphere test), so this only matters if B calls `project` on a point way off the centreline.
- Segment param blending lags the march by one 0.5u step. Deterministic, sub-step, harmless.

---

## → Lane F

1. **Pass `assets` into `createWorld`.** The game currently boots my world but **D's textures
   are loaded and then unused** — I verified every material comes up `USE_DETAIL`-less under
   `boot.js`. In `pickWorld()`:
   ```js
   const mod = await import('./world/index.js');
   const level = await loadLevel(flags.lvl);
   const assets = await (await import('./core/assets.js')).createAssets({ flags, renderer });
   return mod.createWorld(level, { rng: createRng(flags.seed ?? level.seed), assets });
   ```
   (`createWorld` already ignores a missing/failed `assets` — the fallback law holds either way.)
2. **Add the world selfcheck to the qa gate** — it has caught two real defects already:
   ```bash
   # 4b. Lane A world math (headless; no three.js needed)
   node web/js/world/spline.js --selfcheck >/dev/null 2>&1 && green "world selfcheck" || red "world selfcheck failed"
   ```
3. **Screenshot tooling traps**, which will bite you at round-end boot-matrix time. The pane
   drives tabs **hidden**: rAF never fires (so the loop isn't running when you shoot) and
   `innerWidth` is **0** (so the canvas is 0×0 and `toDataURL()` returns `"data:,"`). The pane
   screenshot still looks right because grabbing it fronts the tab. `DBG.shot()`'s anchor-click
   download also silently no-ops here. My harness works around all three in `DBG.size/pose/poseAt`
   — worth lifting into `boot.js`'s DBG. I got PNGs out by having the page `fetch(PUT)` them to
   a 30-line local sink (in my scratchpad, happy to promote it to `tools/` if you want it).
4. **Dev port**: 8140 was already held by another session's server, and `launch.json` is yours,
   so I left it alone and reused the running one. If lanes are meant to run concurrently on one
   box this needs a call.
5. Contract freeze: I'd like `project(pos, hint)` and `flowPulse/hash/stats/dispose` folded into
   TECH.md's world contract, and `assets.texture()` added to the asset contract (see → Lane D).

## → Lane B

- **`wallRho` shrank**: it subtracts `wave.amp + wave.breathe + 0.6` and esophagus `wave.amp`
  is now 1.4 (was 0.9). Playable radius in a base-10 tube ≈ **7.85**. Re-check your graze
  distances. It's time-independent on purpose — a collision surface that breathed would have
  you fighting the wall.
- **`project(pos, hint)`** — pass your previous `s` as `hint` and it's near-free and exact. With
  no hint it seeds from `-pos.z` and still converges (6 fixed-point steps), but the hint is
  strictly better in your per-frame path. Both are asserted in the selfcheck.
- **Crest speed == `biomeAt(s).flow`**, exactly, by construction. `world.flowPulse(s, t)` is
  the *same* function the vertex shader runs (0 = trough, 1 = crest), so pulse ≈ 1 means the
  wall is at its narrowest AND the crest is moving at flow speed right there. Boost window.
- `biomeAt(s).flow` is **blended across segment joins**, not stepped — safe to read per-frame.
- Arena mode: `modeAt(s)` / `arenaAt(s)` → `{center, radius}` for your 6DOF clamp. `collide()`
  branches automatically (sphere test inside a room, radius test in a tube); you don't have to.
- `sample()` allocates fresh Vector3s per call, same as the stub. If that shows up in your
  profile, ask and I'll add `sampleInto(s, out)`.

## → Lane C

- Your schema v1 reads clean here — v1 is additive and my reader consumes L1/L2/L3 unchanged.
  Your ratified arena semantics (sphere spanning `[at-radius, at+radius]`) are exactly what's
  implemented, and `modeAt`/`arenaAt` agree with your notes on all three L3 arenas.
- **Biome ids are the registry in `world/biomes.js`**: `oral · esophagus · stomach ·
  small_intestine · large_intestine · appendix`. Unknown id ⇒ neutral fallback + one warning,
  never a crash. Thanks for writing the "biome = tissue family, not anatomy" rule into TECH —
  that's exactly right and it's why `name` carries the anatomy.
- `curviness` behaves as you documented (0 straight … 1 writhing). Under the hood 1.0 now means
  "as curvy as this radius can safely be", so **you can't author a pinch** — L2 measures 3.32×
  clear, L3 1.89×. Author freely.
- Free lever if you want it: `flow` is per-segment and it now sets the peristalsis wavelength
  (`k = 3.08/flow`). Low flow ⇒ tighter, slower rings. A "sluggish" segment reads sluggish.
- Your L3 note that a sphere is the wrong shape for the stomach body: agreed, see limitations.

## → Lane D

- **Naming mismatch, silent by construction.** Your keys are `wall_smallint_a` / (future)
  colon; my biome ids are `small_intestine` / `large_intestine`. `wall_${biomeId}_a` therefore
  misses, and because assets are optional it falls back to procedural **forever with no error**.
  I've added an explicit slug map in `index.js` so it works today. Proposal for round 2:
  standardise on the biome ids (`wall_small_intestine_a`) and delete the map — one name, one
  place. Your call, just let's not leave it implicit.
- **Your `tile` semantics and mine converged independently** — `[repeats around theta, units of
  s per repeat]`, both read off the stub's uv. Ratified from my side; `[4,16]` looks right in
  engine (`round1_L2_textured.png`).
- **`texture()` is the useful entry point, not `get()`.** TECH names `get(category,name)` as THE
  contract, but it returns raw JSON a shader can't eat. I prefer `assets.texture(name)` and keep
  `get` duck-typed as the documented floor. Suggest F promotes `texture()` into TECH.
- FYI: tiling is applied **in-shader** from `tile`. A raw ShaderMaterial has no `uvTransform`, so
  the `repeat` you pre-apply on the texture is inert here. Not a problem — just don't count on it.
- `wall_esophagus_a` looks genuinely good in engine. The `_b` variants and matcap aren't wired
  yet (round 2, with the TBN work).

## → Lane E

- `world.flowPulse(s, t)` (0..1, crest = 1) is yours to pulse audio/HUD with; it's exact, cheap,
  and matches what the wall is visibly doing.
- `world.biomeAt(s)` → `{id, palette, fog, flow, coatDrain}`. `palette.tint/rim/void` are the
  ART_BIBLE hexes — use `rim` if you want HUD accents that match the biome you're in.

## Round 2 (mine)

Triplanar arena shells + acid plane · villi band (InstancedMesh, small intestine) · normal-map
TBN + matcap · sphincter joint geometry at biome seams · `sampleInto` if B needs it · delete
`_fixture.js`/`dev.html` once boot + C's levels cover them.