feat: dataset creation ui

[Scr4tch587]

What changed

New create view (`components/create-dataset.tsx`, reachable via the "new dataset" header button):

• Proposer section (required): name, Wat Street team, discord user, and a "what is this dataset for?" description — all shown to reviewers in the PR body
• Basics: name/version inputs, calendar + granularity selects, start-date
• Schema: dynamic key/type rows (str · int · float · bool)
• Dependencies: picker populated from `GET /datasets` plus optional lookback (`5d`-style, zod-validated)
• Builder script: CodeMirror editor with Python highlighting and a template pre-filled
• Extras: optional requirements.txt; env-vars checkbox reveals a `.env.template` field with a "secrets never go in the PR" note
• zod + react-hook-form validation mirrors the server's rules (name regex, semver, date format, lookback format), and server-side 400/409 messages render inline on submit
• success screen links the opened PR and repeats the manual-`.env` warning for env-vars datasets

The view is lazy-loaded (`React.lazy`) so CodeMirror's chunk (~540 kB) never loads during normal browsing. shadcn `select`/`checkbox`/`textarea` components added; new deps: `@uiw/react-codemirror`, `@codemirror/lang-python`.

Why

The UI half of the PR-based dataset-creation flow — pairs with the `POST /datasets` endpoint below this PR in the stack.

Benefit

Anyone on the team can propose a dataset without touching git; reviewers get a fully-attributed, pre-validated, pre-linted PR.

🤖 Generated with Claude Code

[github-actions]

• main
  • feat: replace svelte frontend with react + tailwind scaffold #375
    • feat: typed api client and api key handling #376
      • feat: dataset browsing ui #377
        • ci: deploy frontend to github pages #378
          • docs: update specs for react migration #379
            • feat: github rest client for dataset proposals #383
              • feat: dataset proposal service and endpoint #384
                • feat: api client support for dataset proposals #385
                  • feat: dataset creation ui 👈

[Blackgaurd]

[AI] schemaFields.key doesn't .trim(), unlike the other text fields (authorName, team, discordUser, description). A trailing space passes validation silently, and the backend will treat "price " as a different TOML key than "price" — builder output would then fail schema validation with a confusing "missing key" error. Suggest trimming this field too.

[Blackgaurd]

should add another (python) comment giving an example return so people have an idea how to structure the function

[Blackgaurd]

[AI] Unchecking "needs env vars" hides the envTemplate textarea (line 624) but doesn't clear its value, and toPayload here sends env_template whenever it's non-empty, regardless of envVars. A user can end up with env-vars = false in the generated config but a .env.template file still added. Suggest clearing envTemplate when envVarsEnabled flips false, or gating this line on values.envVars too.

[Blackgaurd]

[AI] The red validation hint text doesn't appear on first-time typing because useForm (create-dataset.tsx:132) doesn't set a mode, so react-hook-form defaults to mode: "onSubmit" — fields only validate after the first submit attempt. Consider mode: "onBlur" (or "onChange") so users get feedback as they fill out the form instead of only after clicking submit once.

[Blackgaurd]

[AI] Reported error message looks truncated/garbled on submit failure (e.g. "builder script fails lint:" errors from ruff). Likely cause: the error <p> (create-dataset.tsx:652-654) doesn't set whitespace-pre-wrap, so ruff's multi-line stdout (one line per violation) collapses into a single run-on line under default HTML whitespace handling — nothing appears to be truncated server-side (backend passes the full message through with no length limit). Might be worth preserving newlines here so multi-line backend errors are readable.

That said, actually failing lint on this form isn't itself a problem — backend CI on the resulting PR will catch any lint errors regardless, and the proposer can revise their code from the PR. This is just about making the in-form error message readable, not about needing to prevent lint failures here.

[Blackgaurd]

[AI] Suggestion: add a "reset" button next to the builder script editor that restores field.value back to DEFAULT_BUILDER. Once someone starts editing/experimenting here it's easy to end up in a broken state with no quick way back to the known-good starting template.

[Blackgaurd]

[AI] Object.fromEntries(values.schemaFields.map((field) => [field.key, field.type])) silently drops earlier rows if two schema fields share the same name (last one wins) — no duplicate-key check in the zod schema or before submit. Worth a .refine() on schemaFields checking all keys are unique, so a fat-fingered duplicate doesn't quietly disappear from the proposal.