PROCITY/docs/LANES/LANE_D_NOTES.md

LANE D — NOTES (measured budgets + clip wishlist)

Written by PROCITY-D, 2026-07-14. Standalone verification via web/citizens_test.html. Reference stack ported from 90sDJsim/web/world/index.html ~405–520 (loadRig/spawnRig/_canon/ _rotOnly/head-bone normalize/upgradeStreetPeople). Measurements on the M3 Ultra dev box.

---

ROUND 8 — shop patronage v0 + weather reaction (→ Lane F wiring below)

The crowd comes alive: streamed peds duck into open shops they pass and re-emerge; at night only the open-late video shop draws visitors; rain thins + shelters the crowd. Default-on for the streamed roster (post-flip), behind ?patronage=0 off-switch. Verified in-shell (real plan shops) + test page. qa.sh --strict GREEN. v1 (?roster=v1) path untouched (no patron fields, golden identity holds).

Patronage (D1) ✅

State machine per streamed ped: walking → (every ~10m, if a nearby OPEN shop is within 18m, a seeded roll) → going (steer to the door) → inside (hidden, seeded 5–20s dwell) → emerge (resumes its footpath walk). Hours-aware via the shop's hours.

• Day-long: peds visibly enter/leave shops of all types (noon: up to ~38 inside near a shopfront, 7 types visited).
• Night: at 22:48 only the video shop (hours[1]≥22) is open → patronage is video-only (onlyVideo:true), and its block stays lively (the R6 night floor). The patronage roll ramps up as the streets empty so the few peds out concentrate at the one open shop.
• Deterministic (chunk-keyed identities unchanged: 150 re-derive from seed), leak-free (0 GPU delta over a 2-walk patronage churn — enter/emerge reuses the existing actor lifecycle), and budget-neutral / positive (inside peds are hidden → draws −1 vs patronage-off at the same spot).

Weather reaction (D2) ✅ — reads Lane B's PROCITY.weather contract (does NOT import weather.js)

sim.setWeather({state,intensity}) each frame. Rain (intensity 0.8): density −56% (in the 40–60% target), walk speed +~14%, and patronage chance up → more peds shelter in shops (measured inside 6→25). Overcast: ~−10%. Clear: v1 behaviour. Seed 20261990 rolls clear (v1-identical); use ?weather=rain to force it.

→ Lane F: wiring in web/index.html (F owns the shell)

1. Shop door points — build once from the plan, feed the sim:

const CH = 64, shopsByChunk = new Map();
for (const s of plan.shops) {
  const l = plan.lots.find(x => x.id === s.lot); if (!l) continue;
  const ry = l.ry || 0, fx = -Math.sin(ry), fz = -Math.cos(ry);   // facade normal → the street
  const x = l.x + fx * (l.d / 2 + 0.6), z = l.z + fz * (l.d / 2 + 0.6);   // door at the shopfront
  const k = citizens.chunkKeyAt(x, z);
  (shopsByChunk.get(k) || shopsByChunk.set(k, []).get(k)).push({ x, z, hours: s.hours });
}
if (!rosterV1) citizens.setShops(shopsByChunk);        // patronage needs door points; inert without
if (params.get('patronage') === '0') citizens.setPatronage(false);   // off-switch

(PATRON_RANGE 18m tolerates door-point imprecision; if a door reads wrong, nudge the +0.6.) 2. Weather — in the street branch each frame: citizens.setWeather(window.PROCITY.weather); (B's contract is always a valid {state,intensity}; F sets {clear,0} when ?weather is off.) 3. setNightLivelyChunks (openLate block) — already wired in F1; unchanged.

Baseline note (flip protocol): patronage default-on may move the flags-off draw baseline a hair — in the direction of fewer draws (inside peds hidden), and the default spawn view has ~0 near-rigs so it's near-nil. If F1's regression trips, re-pin per the flip protocol (it's an improvement, not a leak).

---

ROUND 7 — MERGE VALIDATED · GO for the roster flip (→ Lane F: flip on this) · escape hatch ready

Verdict: GO. Lane E's R7 ped-merge closed the one gap the R6 memo named. Full-density streamed roster (perChunk 16) is now under budget in a real town. Flip it default-on. qa.sh --strict GREEN. All numbers in-shell (real town "Boolarra Heads", seed 20261990).

Merge validation (E1 — pipeline/merge_ped.py, 447188a) ✅

• Every one of the 19 peds is now 1 mesh / 1 material (was 8 meshes / 2 materials) → ~1 draw per near-rig; fleet 116→24 draws at the 24-cap (E's own count: 92→19). Bonus: GPU memory dropped too (geometries 205→160, textures 141→90 in-shell).
• Rigs bind + animate (walk/idle) — mixer 0.2–0.3 ms; no T-pose. Skinning/skeleton preserved (65 joints), clips retarget unchanged.
• Silhouettes + atlas materials intact — hi-vis worker (vest/helmet/boots), suit, business, comical kid all read crisp, no atlas seams/melting; identity variety preserved (pickRig unchanged).
• Impostor bake clean (19 subjects, 16×5 atlas); determinism holds (507 chunk-keyed identities re-derive from seed).

The flip numbers (full density, perChunk 16 — no reduced-density concession)

metric	pre-merge (R6)	post-merge (R7)	budget
worst continuous-walk street view draws	356 ✗	241 ✅	≤300
worst view tris	~66k	65,899 ✅	≤200k
worst-view near-rigs	(throttled to 8)	17 (full)	cap 24
draws / near-rig	~7	~1	—
30-chunk soak (3 walks = 726 builds/disposes)	leak-free	leak-free, 0 GPU delta	baseline
heap over soak	50→77→54	78→63 (GC, stable)	stable

Full-density shipping defaults (set this round)

enableStream default is now perChunk 16, radius 2 (was 8). Near-cap stays 24 (now 24 draws, not 168). NIGHT_LIVELY_FLOOR 0.5 for the open-late block (R6). No other knobs change.

→ Lane F: the flip wiring in web/index.html (F owns the shell — decision #1, flip on my word)

Invert the R5/R6 flag: stream is the default; ?roster=v1 is the escape hatch.

const rosterV1 = params.get('roster') === 'v1';   // escape hatch → the old fixed roster
const citizens = new CitizenSim({ renderer, scene, camera, citySeed: plan.citySeed,
                                  graph: plan.streets, fleet,
                                  chunkStream: rosterV1 ? null : { radius: 2 } });  // default-on

Hours-aware (optional, recommended): after construction, citizens.setNightLivelyChunks(keys) where keys = the openLate shop's lot→chunk + 1-ring (sim.chunkKeyAt(x,z); openLate = hours[1] >= 22). Everything else (setTimeOfDay/setExposure/update/setPaused) is unchanged. The hook-driven window (onChunkBuilt/onChunkDisposed → sim.onChunkBuilt/onChunkDisposed) is still the preferred precise option; poll works with zero wiring.

Verified in my harness (test page + runtime-enabled shell): no-flag boot → streamMode:true, perChunk 16, determinism ✓; ?roster=v1 → streamMode:false, fixed roster, determinism ✓. The in-shell no-flag-URL boot completes once F lands the two-line flip above — the roster itself is proven to run correctly in the shell at the shipping default (the soak + flip numbers above are in-shell).

---

ROUND 6 — streamed-roster "default-on readiness" memo (→ Fable + Lane F, for the R7 flip call)

Verdict: functionally ready (deterministic, leak-free, hours-aware, composes) — but full-density default-on is gated on one optimisation (ped sub-mesh merge). Ships budget-safe TODAY at reduced density. Flag stays default-off this round. qa.sh --strict GREEN (sim.js doesn't touch its gates). All numbers below are in-shell (real town "Boolarra Heads", seed 20261990), stream runtime-enabled (F wires ?roster=stream in F2 — call-site in the R5 section below).

Soak (gate-2 + far-field, ?roster=stream) ✅

• Leak-free: 3 full street-graph walks = 242 chunk builds / 242 disposes → 0 GPU geometry / 0 texture delta (renderer.info constant 205/141); 16 shop enter/exit cycles → 0 delta. JS heap oscillates 50→77→54 MB (GC recovers — not monotonic, no leak).
• Determinism: 188 live chunk-keyed identities re-derive byte-for-byte from the seed, in-shell.
• Flag-off byte-identical: shell citizen-0 signature 0:13:7:1.786:1.493:3:-1 (R4 golden), still exact after this round's changes. 0 console errors across the soak.

Hours-aware density ✅ (implemented this round)

Each chunk thins per-frame by the same densityAt(tod) curve v1 uses. New: setNightLivelyChunks(keys) gives the open-late block a night floor (NIGHT_LIVELY_FLOOR=0.5). Measured: the open-late video shop (hours 11–23) block holds ~half its crowd at night while ordinary streets go near-empty (ordinary 65→6, ~90% drop; open-late block 31→16). Lane F call-site: compute the openLate shop's lot→chunk (+ 1-ring) and citizens.setNightLivelyChunks(keys) once after enabling stream (sim.chunkKeyAt(x,z) gives the key). Empty set = uniform thinning (safe default).

Composition with ?dig=1 ✅

The streamed roster is street-tier only: the shell frame loop calls citizens.update in the street branch only, so the roster is fully inert while a shop/dig is open — no shared state, no fight. Proven: enter/exit record shops (the dig context) with stream on is leak-free and the roster resumes intact. (The riffle itself is Lane C+F's flag; couldn't trigger its raycast open via synthetic input without pointer-lock, but the composition claim is mode-separation, which holds by construction.) Same reasoning covers the all-on combo — roster shares no state with winmap (B glass) or dig (C).

Perf / budget — the ONE gap ⚠️ (default-on blocker at full density)

• Stream vs v1 (pop 140) at the busiest node: frame +0.29 ms (0.89 vs 0.60 ms), draws ~parity; most of the ~500 active are far = cheap position-advance only. Mixer stays 0.1–0.3 ms.

• Draw budget: worst continuous-walk street view must stay ≤300 (CITY_SPEC). Root cause: each decimated ped is 8 sub-meshes but only 2 materials → ~7 draws per near-rig. So a dense street of near-rigs blows the budget:

  stream perChunk	worst-view draws	worst near	density (within 70m, median)
  16	356 ✗	15	~26
  12	314 ✗	10	~18
  8 (new default)	291 ✓	10	~13

  So it ships budget-safe at perChunk=8 today (~13 median within 70m — constant + camera-following vs v1's uniform ~13 that thins as the town scales; and it concentrates ~10 near-tier rigs vs v1's 0–1 — the visible upgrade). Full density (perChunk 16, ~2× the crowd) needs the budget headroom.

Recommendation for the R7 default-on flip

1. Land the ped sub-mesh merge (merge each ped's 8 sub-meshes by material → 2 → ~2 draws/rig). Then perChunk 16 worst-view ≈ 260 draws — full density under budget. It's a rigs.js fleet-load optimisation and it is F1-safe in practice: v1's default/spawn view has ~0 near-rigs, so the flags-off draw snapshot is unaffected — the merge only cuts draws where near-rigs cluster (the stream path). Also helps v1 street draws + interior keepers. ~1 session; risk = skinned-mesh merge (verify animation). This is the single thing standing between stream and default-on.
2. Until then: default-on is safe at perChunk=8 (reduced density) — flip is a judgement call on whether ~10 near-rigs + constant far-field beats v1 today, or wait for the merge for the full crowd.

Known non-blockers: near↔mid is a hard LOD switch (documented, imperceptible at 25m); open-shop spawn-bias beyond the night floor is v2-later (needs per-chunk shop data — the sim is graph-only).

---

ROUND 5 — chunk-streamed roster (v2, behind ?roster=stream) ✅ → Lane F wiring below

As-built implementation of the R3 design note (further down). Default-off; v1 path byte-identical. tools/qa.sh --strict GREEN (sim.js doesn't touch the citygen/manifest gates).

What it does. v1 spreads one fixed roster over the whole town, so big-town streets are uniformly sparse. Stream generates + ticks citizens per 64m chunk, keyed so identity is a pure function of (seed, chunkKey, i) — independent of town size / visit order — and windows them to the live chunks around the camera. Density becomes constant per unit street and follows the player. Everything downstream (LOD tiers, 24 near-cap, mixer stagger, rig pool, impostor layer) is unchanged and stays global across live chunks.

Measured (test page, 9×9 fixture grid, ~±208m):

• Constant far-field density: citizens within 70m ≈ 58 at spawn, 129 at 150m out, 83 elsewhere — vs v1 (200 pop) which is uniformly ~12–15 everywhere (spread too thin). Stream is ~5–10× the street density at equal total cost, and it doesn't thin as you walk arbitrarily far.
• Deterministic: 507 chunk-keyed identities re-derive byte-for-byte from the seed (determinism button, stream mode). Same seed + same walk → same residents.
• Leak-free: after warming ped GPU uploads, 2 full grid walks (~380 chunk build/dispose cycles) → 0 geometry / 0 texture delta; rig pool capped at 30, live chunks bounded at 25 (R=2 → 5×5).
• Budget: worst view uses the same near-cap (24) + 1 impostor draw as v1; frame 2.31 ms (~432 fps headroom) with ~500 active (most far = cheap position-advance only). Citizen draw/tri contribution is identical to v1's bound.
• Flag-off = v1: shell with no flag → streamMode:false, citizen 0 signature 0:13:7:1.786:1.493:3:-1 — exact match to the R4 golden capture. No perturbation.

→ Lane F: wiring ?roster=stream in web/index.html (F owns the shell)

Minimum (poll-driven, zero Lane B changes — works today):

const streamOpt = params.get('roster') === 'stream' ? { radius: 2, perChunk: 16 } : null;
const citizens = new CitizenSim({ renderer, scene, camera, citySeed: plan.citySeed,
                                  graph: plan.streets, fleet, chunkStream: streamOpt });

Preferred (hook-driven — the sub-roster window then matches Lane B's built-chunk window exactly; the sim auto-stops polling on the first hook call). Wire onto the chunks ctx only when the flag is on:

if (streamOpt) {
  chunks.ctx.onChunkBuilt    = (key) => citizens.onChunkBuilt(key);     // Lane B calls this (chunks.js:41)
  chunks.ctx.onChunkDisposed = (key) => citizens.onChunkDisposed(key);  // chunks.js:47
}

(chunks already exposes ctx; if not, expose it or pass the two callbacks into createWorld.) Nothing else changes — setTimeOfDay/setExposure/update/setPaused are identical. Flag-off must not construct with chunkStream, so the v1 path (golden identities) is untouched. The 64m chunk key (${cx},${cz}, Math.floor(v/64)) matches Lane B's planutil.js exactly, so B's keys line up with ours.

Hours-aware density is in (task 3): each chunk is thinned per-frame by the same densityAt(tod) curve v1 uses, so streamed streets empty at night too. Open-shop spawn-bias weighting is left for v2-later (needs shop data in the chunk; the sim is graph-only today) — noted, not built.

---

ROUND 4 — in-shell verify + decimation validation (→ Fable, → Lane F)

All three R4 tasks done. tools/qa.sh --strict GREEN. Sign-off for Lane F below.

D1 — in-shell rig/keeper verification (web/index.html) ✅

Verified in the real game (not the test page), via window.PROCITY:

• Placeholder→rig upgrade fires for both peds (citizens.mode → 'rig', 19 fleet, pool active) and keepers (keeper spawns at the counter keeperStand pose as a rig).
• Leak-free: baseline renderer.info.memory = {geometries 213, textures 198}; after 12 enter/exit shop cycles it returned to exactly 213/198 (Δ geom 0, Δ tex 0). Interiors peak ~303 geoms inside, drop back every exit. Confirms the R-review dispose fix (free the keeper's Skeleton only, never the fleet-shared geo/mats) holds in-shell alongside Lane C's interior dispose.
• Determinism in-shell: seed 20261990 → byte-identical identity signature across a full reload.
• ?noassets=1: fleet===null, mode stays placeholder, peds + keeper are placeholder actors, zero ped GLB fetches — town fully playable asset-free.

D2 — decimated ped validation (critical path — E1 landed 518678d) ✅ SIGN-OFF FOR F

Validated Lane E's decimated fleet (web/models/peds/, ref lane-e/round4-peds):

• Per-ped tris: 1564–2805, avg 2450 — all ≤3k (matches E's _peds_decim.json).
• Rigs bind + animate: skinning preserved, the shared walk.glb/idle.glb retarget unchanged (skeleton/65 joints intact), near actors animate, mixer 0.1–0.3 ms.
• No skinning explosions: 24-teleport pool-eviction churn → 0 broken meshes; silhouettes read at street distance; identity variety intact (verified visually — kid/luchador/hi-vis/elder/ business all distinct, no melted hands/faces).
• Impostor atlas bakes clean from the decimated rigs; determinism + leak checks re-pass.
• Gate-3 number for Lane F: at the 24-rig near cap the fleet now contributes ≈59.6k tris (was ~1.2M). Test-page whole-view total 88.9k tris (was ~1.5M). In-shell sample near rig 2729 tris, keeper 2073 tris (was 41k). F is clear to re-measure gate 3 — the fleet leaves ~140k for the town, comfortably ≤200k. (The R1 "peds too heavy" finding is now resolved by E's decimation.)

D3 — exposure sync at dusk/night — FOUND + FIXED A REAL BUG ✅

Verifying the shell exposed a genuine rendering bug (only correct on my direct-canvas test page): the impostor material was toneMapped:false and applied ACES in-shader, but the shell renders through an EffectComposer (RenderPass → bloom → OutputPass). three only tone-maps materials when rendering to the canvas (three.module.js:6921–6931, currentRenderTarget === null); into a composer target, materials output linear and OutputPass applies ACES once. So my self-tone-mapping impostors were double-tone-mapped in the actual game (too dark), matching the rigs only on the test page. Fix (impostor.js): the impostor is now a normal toneMapped:true material — it outputs linear and uses three's own <tonemapping_fragment> + <colorspace_fragment> call-sites (the *_pars_* are auto-injected — including them ourselves double-defines RRTAndODTFit and fails to compile), so it tone-maps identically to the near rigs on both paths: canvas → ACES per-material; composer → linear → OutputPass ACES. Verified: near rig + mid impostor match brightness at noon AND the scene reads correctly at night 22:00 in the shell (docs/shots/laneD/r4_shell_*.jpg), and the test-page canvas path still matches.

• Note for Lane F: citizens.setExposure() is now a no-op (exposure is global via renderer.toneMappingExposure, which both three and OutputPass read). The call at web/index.html:245 is harmless — keep or drop it, your call.

(Round-1/3 notes below remain valid; the tri-budget finding #1 is superseded by D2's decimation.)

What shipped

file	role
web/js/citizens/rigs.js	ported rig stack: fleet loader, _canon/canonRig/_rotOnly, buildFigure (head-bone height-normalise + feet plant), spawnRig (single clip), makeActor (walk↔idle crossfade), pickRig
web/js/citizens/placeholder.js	seeded low-poly box humanoid (POLY lock) — walks/idles, planted at y=0, hot-swaps to a rig
web/js/citizens/impostor.js	4-yaw sprite-atlas baker + instanced billboard layer (mid tier, 1 draw call)
web/js/citizens/sim.js	deterministic roster, footpath lanes, near/mid/far LOD, rig pool, staggered mixer budget, time-of-day density
web/js/citizens/keepers.js	one keeper per shop at the counter slot, idle + greet head/body-turn
web/citizens_test.html	standalone harness: fixture 3×3 street graph, sliders, tier debug, determinism check
web/models/peds/*	19 rigged GLBs + walk.glb + idle.glb, byte-identical copies from 90sDJsim (checksummed)

The fleet (copied, never edited)

19 rigged peds (17 normal + 2 comical) + 2 clip-only GLBs, ~37 MB total, copied from johnking@100.91.239.7:~/Documents/90sDJsim/web/world/models/peds/ (also present locally at ~/Documents/90sDJsim/…). SHA-verified byte-identical (house law: canonical source stays upstream).

• Skeletons canonicalise cleanly: mixamorig1…→mixamorig…, so one shared walk clip drives all 19 and one shared idle clip too. Verified in-scene — a single walk.glb animates every ped.
• walk.glb/idle.glb position tracks + Hips.quaternion are stripped (_rotOnly) — without this the different-scale root track inflates peds to giants. Confirmed: no giants/ants across all 19.

Measured budgets (M3 Ultra, citizens_test.html)

Preview note: the in-app browser reports visibilityState:'hidden', which throttles requestAnimationFrame, so the on-screen fps counter reads low (47–84). True cost was measured by timing 140 manual update()+render() frames.

200 citizens, midday (the acceptance fixture):

metric	measured	CITY_SPEC / Lane-D budget	verdict
near (rigged) actives	24 (hard cap)	≤ 24	✅
mid (impostor)	~122	—	—
far (culled)	~54	—	—
mixer update / frame	0.4 ms max, ~0.2 ms avg	< 2 ms	✅ comfortably
sim logic / frame	0.24 ms avg	4 ms build budget	✅
full frame CPU (update+render)	~2.6 ms → ~380 fps headroom	60 fps (16.6 ms)	✅ huge margin
mid tier draw calls	1 (69 instances, 138 tris)	1 per atlas	✅
impostor atlas texture	2048×640 (16×5 cells, 4 yaws × 19 peds)	≤ 2048, < 512 MB total	✅ (~5 MB)
determinism	live roster ≡ seed recompute, stable while walking	same seed → same crowd	✅

Time-of-day density (200 base): 00:00 → 12 active, 06:00 → 70, 12:00 → 200, 15:00 → 150, 21:00 → 36. Curve in sim.js DAY_CURVE.

⚠️ Findings for Lane E / Lane F (need a decision)

1. The inherited peds are ~49.7k tris each — too heavy for the CITY_SPEC 200k-tri "typical view". 24 near rigs ≈ 1.2M tris; a realistic 16-rig street view already measured 1.02M tris. Framerate holds on M-series (GPU eats it), but this leaves nothing for Lane B buildings. Recommendation: decimate the ped fleet to ~5–8k tris in the asset pipeline (Lane E) — at 6k/ped the near cap costs ~144k tris, back within budget. The rig stack is indifferent to poly count; this is purely an asset job. Until then, NEAR_MAX (sim.js) is the throttle.

2. Draw calls scale with near rigs (~6 calls/ped — each GLB has several sub-meshes/materials). 16 rigs ≈ 144 calls; 24 rigs ≈ ~190. Under the 300-call street budget today, but tight once Lane B adds shells. Recommendation: merge each ped's sub-meshes by material on load (BufferGeometryUtils.mergeGeometries) → ~1–2 calls/ped. Cheap win, deferred (not blocking).

3. Peds are metallic-PBR (metalness≈0.5) and render black without a scene.environment. The test scene builds a neutral PMREM; the impostor bake is fed the same env. Lane B's shell must set scene.environment (a sky PMREM) or every citizen — near and mid — goes dark. This is wired: CitizenSim reads scene.environment and passes it to the impostor baker.

4. Impostor material is toneMapped:false and does ACES + sRGB itself (matching ACESFilmicToneMapping, exposure /0.6). The shell DOES animate renderer.toneMappingExposure (Lane B lighting.js sets it per day segment), so mid billboards drift brighter/darker than the near rigs across the day unless matched. Round-3 fix: CitizenSim.setExposure(e) passthrough added — see the Lane F hook below. (For a non-ACES curve the in-shader ACES in impostor.js IMP_FRAG would also need to match; today both are ACESFilmic so exposure is the only variable.)

Hardened by a 4-dimension adversarial code review (see D-progress.md): 6 confirmed defects fixed — non-deterministic fleet order, shared-geometry disposal, impostor under-exposure, unrestored viewport, missing modelMatrix, frozen exposure.

Integration hooks for Lane F

• new CitizenSim({ renderer, scene, camera, citySeed, graph, fleet }) — graph = { nodes:[{id,x,z}], edges:[{id,a,b,width,kind}] }. Swap the fixture graph for Lane A's CityPlan.streets. Footpath lanes are derived from edge width + a 0.9 m margin; pedestrians keep to the right of travel.
• sim.setPopulation(n) (slider / density) · sim.setTimeOfDay(t01) (drive from the shell's day segment) · sim.setPaused(bool) (wire to visibilitychange) · sim.update(dt) each frame.
• [Round-3] sim.setExposure(e) — one line, needed for dusk/night. The shell already calls citizens.setTimeOfDay(...) in the street loop (index.html:225); add right after it:

  citizens.setExposure(renderer.toneMappingExposure);   // [Lane F] keep mid billboards matched to day/night exposure
  

  Cheap (sets one uniform), safe to call every frame, and survives the placeholder→rig atlas re-bake (stored internally). Verified in citizens_test.html: uniform tracks 0.55↔2.3 and persists across the fleet upgrade. Without it, mid impostors read too bright at night / too dark at noon vs the near rigs. (This is a Lane-F seam edit in index.html; flagged here per the cross-lane note rule.)
• [Round-3] Keeper fleet-upgrade (enables F §3.4). Keepers today are placeholder-only because F builds new KeeperManager({ camera, citySeed }) with no fleet. To upgrade them to shared GLB rigs, pass the loaded fleet: new KeeperManager({ camera, citySeed, fleet }). keepers.js already picks a rig when fleet.ready and falls back to a placeholder otherwise — no other change. Dispose is leak-free: verified 15 enter/exit cycles with a rig keeper → renderer.info.memory geo/tex return exactly to baseline (_disposeInner frees the clone's skeleton bone-texture only; shared fleet geo/mats are preserved for siblings). Safe to spawn/remove() per interior enter/exit.
• Chunk streaming (v1.5): roster is currently whole-graph — fine for v1, but big towns look sparse (see the design note "Chunk-streamed roster" below for the why + the full plan).
• Keepers: keepers.spawn(roomGroup, { x, z, ry, shopId, type }) — feed x,z,ry from Lane C's interior counter places; call on interior build, keepers.remove(handle) on dispose. keepers.update(dt, playerPos) each interior frame.
• ?noassets=1 verified: full placeholder town, mixers=0, zero crashes.

Chunk-streamed roster — design note (✅ IMPLEMENTED in R5, behind ?roster=stream)

Round-3 task 3 (design), built in Round 5 — see the "ROUND 5" section at the top for the as-built result + Lane F wiring. This section is the original spec; the implementation followed it: hook-driven (onChunkBuilt/onChunkDisposed) with a poll fallback, chunk-local identity, global near-cap, owner despawn, constructor opt-in (chunkStream). Kept for the rationale.

The problem (measured). setPopulation(N) builds one global roster and spreads it uniformly over all edges (identityOf picks edge = rng()*edgeCount across the whole graph). On the 12-edge test fixture that fills the view. On a real town — "Boolarra Heads" has hundreds of edges — those same N citizens scatter town-wide, so almost all sit beyond the 70 m cull from any one camera (I measured a 27-node generated town: pop 200 → 4 near / 6 mid / 190 far). Raising N to fill the near streets wastes roster + advance cost on citizens nobody can see, and still can't guarantee local density.

The fix. Generate + tick citizens per chunk, keyed so identity is independent of town size and visit order, and stream them in lockstep with Lane B's chunk window (the chunks already building/ disposing around the player). Density becomes constant per unit street, and total roster is bounded by the live-chunk window, not the town.

Design:

1. Chunk-local identity. Replace the global running id with a per-chunk key: identityOf(citySeed, chunkKey, i) for i in [0, perChunkCount), and choose the citizen's home edge from that chunk's edges (chunkIndex(plan).chunks[chunkKey].edges), not the global list. Then "same seed + same chunk → same people" holds regardless of what else exists or the order chunks were visited — the determinism property that whole-graph keying quietly loses at scale. perChunkCount is seeded per chunk and scaled by district (main-street chunks busier than residential) × densityAt(tod).
2. Windowing — poll-driven (works today, no Lane B change). Each frame derive the active chunk set from the camera: chunkKey(⌊camX/chunkSize⌋, ⌊camZ/chunkSize⌋) + neighbours within a radius R (R≥1 so a walker never steps into an unloaded chunk before its owner unloads). Diff vs the live set → build sub-rosters for newly-active chunks, dispose for newly-inactive. This mirrors B's own window and needs nothing from B. Alternative — hook-driven: chunks.js exposes an optional ctx.onChunkBuilt (see chunks.js:40 — "inert unless a consumer sets it"); if B also adds onChunkDisposed, drive spawn/despawn off those instead of polling. Prefer this once B commits the pair; poll until then.
3. Storage. Swap the single this.roster array for this.chunkRosters = Map<chunkKey, Citizen[]>. The per-frame update() iterates the union of live-chunk citizens. Everything downstream is unchanged — LOD tiers, 24 near-cap, mixer stagger, rig pool, impostor layer already operate per-citizen and are chunk-agnostic; the near-cap + mixer budget stay global across live chunks (nearest-first selection already does the right thing on the merged set).
4. Ownership + hand-off. A citizen is owned by its spawn chunk. It walks freely (the graph is continuous); because R≥1 keeps neighbours loaded, crossing a chunk boundary is seamless. It despawns only when its owner chunk unloads — even if it's momentarily standing in a still-loaded neighbour (acceptable: that's behind the player, past the cull). No re-keying on cross, so no identity churn.
5. Budget. Live window ≈ (2R+1)² chunks × perChunkCount. Tune perChunkCount (~15–30) so the near streets reach the 24-cap while total roster stays ~150–270 — similar cost to today's flat 200, but concentrated where the camera is instead of smeared across the map.

Migration. Opt-in and back-compatible: new CitizenSim({ …, chunkStream: { chunkIndex, chunkSize, radius } }) switches on chunk-keyed identity + windowing; omit it and the current whole-graph roster is unchanged. So v1 ships as-is and v1.5 flips a constructor option — no rewrite of the hot path.

Explicitly NOT in scope here: cross-chunk social groups, per-district behaviour trees, persistent citizens (a citizen you saw yesterday) — all v2 (docs/V2_IDEAS.md).

Clip wishlist (for the mixamo-fetch run — CHECK THE 34 EXISTING CLIPS FIRST)

Only walk + idle are wired today (they ship with the ped fleet). CITY_SPEC says ~/Documents/mixamo-fetch/out/ already holds 34 clips (sit, lean, look-around, phone, …) — these bind directly through the canonical mixamorig skeleton exactly like walk/idle, so wiring them is a data task, not new rig work. Verify each exists before requesting; never re-download.

Wanted for richer loiter/keeper behaviour, in priority order:

1. sit — bench-sit at benches/verandah steps (sim loiter already has the hook; needs the clip).
2. lean — loiter against a shopfront (window-shopping stops).
3. look-around / idle-look — window-shopping variety + a better keeper greet than the body-turn.
4. talk / gesture — pairs chatting on the footpath (spawn in seeded 2-groups).
5. carry / laden walk — shoppers with bags after a purchase (content-phase tie-in).
6. browse / reach — keeper stocking shelves; customer reaching a shelf inside interiors.

Bespoke (won't be on Mixamo, hand-author later): crate-riffle at market stalls, till-operation.

Known limitations (honest)

• Near↔mid swap is a hard LOD switch at ~25 m (hysteresis 24/27 m). At that range a 128px impostor and the full rig subtend near-identical screen size, so it reads as seamless — but it is a switch, not a cross-fade (rig materials are shared across SkeletonUtils clones, so per-instance opacity fade isn't free). If a pop is ever visible after buildings land, the fix is a short dither/fade band.
• In ?noassets mode, mid impostors use 8 generic placeholder variants while near placeholders are per-citizen coloured — so a citizen's colour can shift slightly crossing 25 m. Fallback-only; with the real fleet, near and mid are the same baked ped.
• Loiter timing is dt-driven (cosmetic), so exact positions at time T aren't reproducible across runs; identity (who, which ped, height, speed, spawn beat) is fully seeded and asserted.