Fable (integrator): round-7 lane instructions — roster default-on flip + GODVERSE real stock + weather + strict harness

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This commit is contained in:
m3ultra 2026-07-14 23:53:54 +10:00
parent 6948372f86
commit 15e2cc0a9f
2 changed files with 231 additions and 0 deletions

View File

@ -1,5 +1,8 @@
# PROCITY — Round 6 lane instructions (from Fable, integrator)
> **⚠ SUPERSEDED by [ROUND7_INSTRUCTIONS.md](ROUND7_INSTRUCTIONS.md) (2026-07-14).** Round 6 is
> complete — ALL six lanes delivered; v2.0-alpha tagged (`6948372`). Historical reference only.
Date: 2026-07-14 · Written after the Fable round-5 review (`d71420f`). You are an Opus 4.8
agent assigned ONE lane (read your session prompt). Read this file, your lane spec, your
`<X>-progress.md`, and NOTES addressed to you before touching code.

View File

@ -0,0 +1,228 @@
# PROCITY — Round 7 lane instructions (from Fable, integrator)
Date: 2026-07-14 · Written after the Fable round-6 review (`6948372`) and the **v2.0-alpha
tag**. You are an Opus 4.8 agent assigned ONE lane (read your session prompt). Read this
file, your lane spec, your `<X>-progress.md`, and NOTES addressed to you first.
Repo state: everything committed on `main` at `6948372` = `v2.0-alpha`. `tools/qa.sh
--strict` GREEN **5/5** (incl. the v2 flags harness at warn level). Four v2 flags live,
default-off: `?plansrc=osm` · `?winmap=1` · `?dig=1` · `?roster=stream`. Goldens: synthetic
`0x3fa36874`, osm `0x34cfdec0`. Interior law ≤350 draws/room (worst 313, sweep asserts).
27 procity GLBs live. fal balance ~$21.36. Tags `v1.0`/`v1.1`/`v2.0-alpha` await John's push.
## Theme: the flip round + the content firehose
Round 6 built the machinery; round 7 turns two keys:
1. **`?roster=stream` becomes the DEFAULT.** D's readiness memo says it's deterministic,
leak-free, hours-aware, and composes — gated only on one optimisation (the ped sub-mesh
merge, so full density fits the draw budget). E merges, D validates, F flips.
2. **Real stock enters the dig** (the reason this system exists — V2_IDEAS #1, deferred
twice): a curated GODVERSE content pack feeds C's `stockAdapter` behind `?stock=real`,
so the crates you riffle hold *real* record sleeves instead of parody canvas.
Plus: the flags harness graduates **warn → strict**, B ships weather, A hardens osm mode.
End-of-round milestone: **`v2.0-beta`** (F tags; criteria in decision #6).
## Prime-law amendment (read carefully — the baseline moves this round)
The v2 prime law ("default boot stays v1-identical") gets its first **deliberate,
versioned** exception: when F flips the roster default, the default boot changes. Protocol
(decision #1): the flip commit and the flags-off-baseline re-pin happen in ONE F-owned
commit, with the old baseline recorded in the commit message. **Plan goldens do NOT move**
(`0x3fa36874` / `0x34cfdec0` — the plan layer is untouched by the roster). Any other lane
whose feature would move the default boot: that's still a land-blocker — the exception is
the flip only.
## Integrator decisions (so nobody blocks)
1. **Roster-flip protocol**: E1 (sub-mesh merge) → D1 (validate + full-density numbers) →
F1 (flip default; keep `?roster=v1` as the escape hatch back to the fixed roster;
re-pin the flags-off draws/tris baseline in the same commit). Order is strict — no lane
pre-empts the flip.
2. **Real-stock scope**: read-only content, NO economy backend. E builds the pack offline
from GODVERSE (`~/Documents/dealgod/GODVERSE.md` — recordgod's 185k vinyl w/ images is
the target): curate ~200400 sleeves per record-shop archetype (+ optional small packs
for book/toy if cheap), bake into atlas textures + a JSON index (id, title, artist,
price band, atlas UV), publish atlases to the depot like any asset. C consumes behind
`?stock=real` through the existing `stockAdapter` seam. Buying stays a stub. `?noassets`
and flag-off behavior untouched. **No runtime network beyond the existing depot path;
no databases; nothing from GODVERSE checked into the repo except the built pack index
if it's small.** Parody-name law: run titles/artists through the same
no-real-trademarks screen as everything else — if the source data is real-world
metadata, parody-transform it in the pack builder (deterministic).
3. **Weather is B's round**, behind `?weather=1`: seeded per-day weather state (clear /
overcast / rain), rain particles + wet-ground response, wind sway on awnings/trees if
cheap. The sky domes for it already exist in `web/assets/gen/`. Tram loop is a stretch
goal ONLY if weather lands early (`?tram=1`, instanced, one path over B's bus_shelter
stops, no traffic sim).
4. **Harness graduates to strict** (F, after the flip lands): a red flags-harness now
fails qa.sh. New flags this round (`stock`, `weather`, `tram` if it lands) enter at
warn and graduate next round.
5. **fal**: no new spends. Broken-prop re-gen clause (≤2 attempts, logged) still stands.
6. **Tag `v2.0-beta`** (F, end of round) when: roster default-on + escape hatch verified ·
flags harness strict + green · `?stock=real` riffles real sleeves in a record shop ·
`?weather=1` rains · qa strict green. If stock OR weather slips, tag anyway with what
landed and note it — same no-hostage rule as always.
## Process rules (unchanged)
- ONE shared working tree. `git status` first; stage only your files by exact path;
options before `--`; `add` new files first; commit immediately; pin a durable lane ref.
- Leave `qa.sh --strict` GREEN. Update `<X>-progress.md`; cross-lane asks in your NOTES.
- F owns `web/index.html` + `tools/`; document call sites for F.
- `procityprogress.txt` / `settings.local.json` stay untracked.
- Repo `/Users/m3ultra/Documents/procity`; GODVERSE reference `~/Documents/dealgod`;
dev server `:8130`.
## Sequencing
```
E1 (ped sub-mesh merge) ──→ D1 (validate + density numbers) ──→ F1 (FLIP + re-pin baseline)
E2 (GODVERSE stock pack) ──→ C1 (?stock=real in the dig) ──────→ F2 (strict harness + new smokes)
B1 (?weather=1) · A1 (osm parity + 2nd town) ── parallel ──────→ F3 (tag v2.0-beta)
```
E has both feeder tasks — start E1 immediately (it gates the flip), E2 right after.
F1 lands before F2 (strict) so the baseline only re-pins once.
---
## Lane A (Citygen) — osm mode to parity + a second town
The Melbourne import works; make osm a first-class citizen, not a demo.
1. **Selfcheck parity**: parameterize the full selfcheck suite by plan source so osm plans
get the same invariant coverage as synthetic (lots/shops integrity, hours contract,
exactly-one openLate, storeys ranges, determinism, <100ms). Wire so `qa.sh` runs both
(F already pins both goldens — extend the checks behind them).
2. **A second fixture town**: extract another AU town from the thriftgod cache (different
scale/shape than inner-Melbourne — a smaller coastal or regional strip is ideal) →
`?plansrc=osm&town=<key>`. Same laws: checked-in module fixture, zero network,
deterministic, pinned golden per town (give F the hashes).
3. **Document the "add a town" recipe** in LANE_A_NOTES (fixture extraction → module →
golden pin) so future towns are mechanical.
4. Watch for: osm towns must satisfy the SAME downstream contracts (openLate exactly-one,
registry types, hours sanity) — if a real town's data fights a contract, the importer
normalizes it (and logs what it changed), the contract doesn't bend.
**Acceptance**: full selfcheck green on every (source, town); second town boots end-to-end;
recipe documented; both plan goldens unchanged, new town's golden pinned; qa green.
## Lane B (Streetscape) — weather
`?weather=1` (decision #3). Seeded, deterministic, cheap.
1. **Weather state**: seeded per (citySeed, day) — clear / overcast / rain (weights feel
AU-plausible). Expose `window.PROCITY.weather` for other lanes (D may thin crowds in
rain later; not this round).
2. **Rain**: a particle layer (instanced/points, ONE draw), audible-later, camera-following;
wet ground = roughness/darkening response on street surfaces; puddle sheen if cheap.
Sky: pick the matching dome from the existing `sky-*.jpg` set (summer storm / grey
drizzle are already in `web/assets/gen/`).
3. **Wind**: subtle sway on awnings + trees during overcast/rain if it fits the budget —
vertex-shader time offset, not per-frame CPU.
4. **Budget + composition**: worst view ≤300 draws with weather on; works with winmap
(rain outside a lit window is the money shot — screenshot it), night, and interiors
(rain audio/visual stops inside; interiors are separate scenes so mostly free).
5. **Stretch ONLY if 14 land early**: `?tram=1` — one instanced tram/bus on a seeded
main-street loop over your bus_shelter stops, door-stop pause at each. No traffic sim.
**Acceptance**: `?weather=1` deterministic per seed/day; rain + wet ground + matching sky
render; ≤300 draws; composes with winmap/night; flag-off identical; screenshots (incl.
rain-through-window); qa green.
## Lane C (Interiors) — real sleeves in the crates
`?stock=real` (decision #2). The stockAdapter seam you built in v1 finally eats.
1. **Define the pack contract with E** (you own the consumer side): what the JSON index +
atlases must provide per item (title/artist/price-band/atlas+UV). Put the agreed schema
in LANE_C_NOTES *first* so E builds against it — a half-day handshake beats a rework.
2. **Consume**: when `?stock=real` and the pack is available (depot or `?localdepot=1`),
record-shop bins draw sleeves from the pack (seeded pick per bin — same seed, same
crate contents) instead of parody canvas; the dig riffle shows real covers + pack
titles/prices; the pull-out card shows title/artist/price.
3. **Laws**: deterministic per shop.seed · leak-free (atlas textures dispose with the room
or share via the loader cache — prove with the soak) · ≤350 draws (atlas = the sleeves
stay batched; assert via drawSweep) · `?noassets` and flag-off untouched · fail-soft
(pack missing → parody canvas, warn once).
4. **Extend the dig soak**: `?dig=1&stock=real` enter/riffle/exit × N rooms, zero leaks,
determinism.
5. If E's pack covers book/toy too, wire those shelves as a bonus — record shops are the
acceptance bar.
**Acceptance**: riffling a bin with `?stock=real` shows real (parody-screened) sleeves with
titles/prices, deterministic + leak-free + ≤350 draws; fail-soft proven; contract doc'd;
qa green.
## Lane D (Citizens) — validate the merge, hand F the flip
1. **D1 — validate E1's merged peds** (the flip gate): rigs bind + animate (walk/idle),
silhouettes/materials intact, impostor bake clean, identity variety preserved. Then the
numbers: draws-per-near-rig before/after, worst-view draws/tris at FULL density
(`?roster=stream`, no reduced-density concession), 30-chunk soak leak-free. Write the
go/no-go into LANE_D_NOTES with the exact figures — F flips on your word (decision #1).
2. **Full-density defaults**: with the merge in, set the stream path's density/near-cap
defaults to the values your memo said full density needs — these become the shipping
defaults when F flips.
3. **Post-flip verification** (after F1 lands): default boot runs the streamed roster —
re-run your determinism + far-field-density + leak checks on a NO-FLAG boot; verify the
`?roster=v1` escape hatch actually restores the fixed roster.
4. Do NOT start rain-behavior/schedules/queues — crowds react to weather in a later round.
**Acceptance**: merged peds validated with published numbers; go/no-go memo written; after
the flip, no-flag boot passes your full check suite; escape hatch verified; qa green.
## Lane E (Assets) — the two feeders
Both critical-path. E1 gates the flip; E2 gates the stock. In that order.
1. **E1 — ped sub-mesh merge (START IMMEDIATELY).** Per-ped GLBs carry multiple
primitives/materials → multiple draws per near rig. Merge to (ideally) ONE primitive +
one atlased material per ped: gltf-transform `join`/`palette`/`instance` pipeline —
**preserving skinning (joints/weights), morphs if any, and compatibility with the shared
walk/idle clips**. Backup originals (gitignored), process all 19, verify each loads +
binds in the vendored GLTFLoader, record before/after draws+tris in AUDIT.md, hand to D.
2. **E2 — the GODVERSE stock pack.** Build `pipeline/build_stock_pack.py`: read recordgod
data from `~/Documents/dealgod` (fixture-extract like A's pattern — the pipeline reads
the local dataset at build time; NOTHING queries it at runtime), curate 200400 sleeves
(image quality + variety), **parody-transform titles/artists deterministically** (the
no-real-trademarks law applies to metadata too), bake atlas PNG/WebPs + JSON index to
C's schema (wait for C's contract note — it's their task 1), publish atlases to the
depot, stage for `?localdepot=1`. Provenance in AUDIT.md (source dataset, count,
transform). Small packs for book spines / toy boxes only if the record pack lands early.
3. **On-call**: B may want a rain-streak sprite or puddle texture (procedural first —
generate only if truly needed, flux_local, free).
**Acceptance**: E1 — all peds single-draw (or documented floor), skinned + animating, D
signs off; E2 — pack published + staged, index validates against C's schema, parody screen
run, provenance logged; qa green (incl. your own provenance-drift gate).
## Lane F (Integration) — flip, tighten, tag
1. **F1 — the roster flip (after D1's go).** In `web/index.html`: streamed roster becomes
default; `?roster=v1` restores the fixed roster; `?roster=stream` stays accepted (now a
no-op). **Same commit**: re-pin the flags-off regression baseline (draws/tris snapshot)
to the new default boot, old values in the commit message (prime-law amendment). Plan
goldens untouched. Then boot no-flags + escape hatch and eyeball both.
2. **F2 — harness to strict (after F1).** Flags harness failures now fail qa.sh. Add
smokes for the new flags as they land: `stock=real` (enter record shop, open dig,
assert non-parody sleeve count > 0), `weather=1` (boot, assert particle layer + wet
response active, 0 errors), `tram=1` if B ships it. New flags enter at warn (decision
#4), existing four + flags-off are strict.
3. **F3 — tag `v2.0-beta`** (decision #6 criteria): summary tag message with the flip
baseline note; update flags table + runbook + F-progress. Local; John pushes (remind
him v1.0/v1.1/v2.0-alpha are also still unpushed).
4. **Marshal**: this round has two hard sequams (E1→D1→F1, E2→C1) — if a lane stalls,
flag it in F-progress early rather than at round end. The R5 lesson (silent no-show)
now has a harness for code, but not for schedule — you're the schedule harness.
**Acceptance**: default boot = streamed roster with re-pinned baseline + working escape
hatch; harness strict + green with new-flag smokes; `v2.0-beta` tagged on a green tree;
runbook/flags table current.
---
*— Fable (integrator). Two keys turn this round: the town gets permanently busier, and the
crates get real. Both have hard sequencing — respect the arrows, commit atomically, pin
your refs, leave it green.*