[ishaanko]

The codebase had a lot of slop: the client was split by layer (components/, pages/, lib/, data/), so one mode lived in four folders; Daily.tsx and ui.tsx were god files; game numbers and validators were copied between the schema, the rules and the page copy; jobs carried v.any() args that the worker API cast back; and some arguments were never checked at the API boundary.

Now the client has a folder per mode (daily/, dex/, bounties/, lab/, map/, plus home/, learn/, players/), with shared neuron widgets in neuron/ and generic primitives in ui/. The backend has one source for each rule, and bad states and bad input can't get in.

• Layer-first folders: src/ is grouped by feature. worker/places.py now writes src/map/zones.json.
• God files: Daily.tsx (486 lines) is split into the page, result card, guess list, heat review and archive. ui.tsx is split into generic ui/ primitives and neuron/ parts.
• Duplication: the effect and rarity validators come from rules.ts. Profile and puzzle fields are declared once and reused by the worker API. Heat bands have one threshold table, used by the bar, chart, map and share grid. The example game reuses the daily's guess row. Lab's Metric became Stat. checkPush and profileOf replace repeated inline code. PageTitle, DoseInput and buttonClass replace copied markup. JoinDialog uses Modal.
• Representable illegal states: rarity is a union, not a string (removes 6 casts). Jobs are a union per kind, so job.kind === "judge" gives typed args. daily.current nests answer inside result, and lab.mine returns result only once one exists, which removes the ! assertions in Daily and Lab.
• Layering and pass-throughs: lib.ts no longer re-exports rules.ts. The client no longer imports convex/bountyPool (where the hidden judge prompts live). lab.ts no longer imports DOSE from bounties.ts. Stats moved out of jobs.ts into stats.ts. lib/ no longer imports components. The session/JoinDialog import cycle is gone.
• Implicit coupling and shotgun surgery: dose, efficiency, bounty length and reply length live in rules.ts, and the copy reads them. About and the field notes read the placed and quiet counts from zones.json. Clerk and the theme-color meta read theme colors from the CSS instead of keeping copies. Nav and footer links sit together.
• Validation at the boundary: malformed dates, search layers and map squares are refused where they come in, and catchNeuron checks its neuron. /daily/<garbage> shows "No such puzzle." Saved field notes are filtered to known ids.
• Worker: the four copy-pasted job loops are now one dispatch, and the sweep cursor path reuses STORE.

Deploy notes

• The schema is stricter: rarity and the job rows. Convex checks existing rows when it deploys. If an old prod row doesn't match, the deploy fails and the current version stays live; no data is touched. The local dev data passed.
• Two query shapes changed (daily.current, lab.mine). A tab left open across the deploy may break on a finished daily until it reloads.
• Worker changes take effect only after a pull and restart of neuronaut-prod-worker. I didn't restart it.

Verification

• bunx tsc -p convex --noEmit, bun run build, bun run lint (20 warnings, down from 23) and bun run test (81) pass. Worker pytest passes (29).
• Pushed to a copy of the local dev backend, then walked every page in Chromium with no page errors. Played practice puzzle fix: reject non-finite bounty steering doses #1 to a solve as a guest; the result card, the solved panel, the answer on the map and "Why each guess scored" all render from the new result.answer.
• Checked the worker dispatch with stubbed handlers: same order as before (judge, par, daily, combo), and failures still go to worker:fail. Not run on the GPU.
• e2e/session/run.mjs fails all 15 cases on main too: its about:blank page has no crypto.randomUUID. With a polyfill added to a local copy, it passes 15/15 on both main and this branch. The full e2e/game.mjs wasn't run, because it needs a second GPU worker next to the prod one.

A follow-up design PR is stacked on this branch.

Claude Opus 5.5 in Claude Code

🤖 Generated with Claude Code

[greptile-apps]

Confidence Score: 5/5

[Medium impact] The PR appears safe to merge; the reported search crash is fixed.

Summary

This PR groups the client by feature, shares backend rules and validators, and tightens input checks.

• The backend checks player inputs against shared rules before using them.
• Each GPU job kind now carries its own argument shape through the worker.
• The daily game shows its answer inside the finished result.
• Client code now lives beside the feature that uses it.
• Map pages and their generated zone data now share a feature folder.

_{Reviews (2) · Last reviewed commit: "fix(web): don't send the lens search a l..." · Reviewed by Greptile}