Simplify the canonical D1 Worker example

[mkuznets]

The canonical example was spending most of its lines on HTTP plumbing rather than on the generated D1 code it exists to demonstrate. Three commits, each independently green, so the mechanical reformat stays out of the way of the substantive changes.

1. Unify formatting on the repository configuration

examples/d1-worker/ and test/miniflare/ each carried byte-identical copies of .prettierrc, .editorconfig and .gitignore, so one code style was maintained in two places and disagreed with the root in all three. They now inherit the repository configuration — 120 columns, spaces, double quotes.

The win beyond deduplication: those directories hold checked-in plugin output, which has always been 2-space and double-quoted. Under the old split, the example demonstrating the generated code looked foreign sitting next to it.

The .editorconfig copies also carried a latent bug. Their [*.yml] section never matched sqlc.yaml, the only file it was meant to protect, so YAML inherited indent_style = tab — illegal for indentation in YAML. Any editorconfig-aware editor would have written a file sqlc cannot parse.

2. Run the fixture sub-projects on npm

Those two directories were the only part of the repository on bun, for no reason the tree records. Everything else already runs on the pinned Node and npm, so they brought a second package manager, a second lockfile format and a pinned oven-sh/setup-bun along with them. bun.lock becomes package-lock.json, the Makefile targets run npm ci / npm run, and the setup action loses its bun input, install step and version echo.

3. Rewrite the canonical Worker on Hono

Manual URL parsing, a path regex, hand-rolled method branching and a request-wide try/catch are gone. What remains is two route declarations, the query calls, and one onError that keeps the only error handling actually about this project — SqlcD1Error means a generated query failed, anything else does not.

No validation layer replaced parseId and isRenameBody, because the generated runtime already validates its own arguments. A malformed id or missing name raises QueryArgumentError before D1 is called:

request	response
GET /users/not-an-id	500 query_failed — QueryArgumentError from the generated code
PATCH {"nickname":"…"}	500 query_failed — same
PATCH with malformed JSON	500 internal_error — Hono's parse error, correctly not labelled a query failure

So the malformed-input test changed character rather than disappearing: it now asserts the generated argument checks firing, which is a demonstration of the library instead of a test of hand-written HTTP plumbing.

Trade-offs, deliberate: bad input answers 500 rather than 400 — a real API would validate at the edge, but this one exists to show the query surface. The 405 responses went too; they needed two fallback routes in a specific registration order, and nothing about D1 is clearer for having them.

src/index.ts goes from 46 lines to 35, with one dependency added.

Verification

make clean && make test from a fresh plugin build: 80 unit, 27 candidate, drift clean, 33 miniflare, 5 example. make fmt-check clean repo-wide. wrangler deploy --dry-run bundles the example at 87.93 KiB with the env.DB binding resolved, which exercises the new dependency through wrangler's bundler rather than only vitest's. Commits 1 and 2 were each checked out into a worktree and verified on their own so git bisect never lands on a broken tree. The four generated fixture files are byte-identical — make fmt never touched plugin output.

One thing left open

npm audit reports seven findings that bun never surfaced — esbuild, sharp, undici and ws, all transitive under miniflare, wrangler and vitest-pool-workers. Dev-only and predating this PR. Those lockfiles are a deliberately pinned test environment, so bumping them is a separate decision rather than something to slip in here.