Migrate to McpServer.registerTool; fix double-encoded SoQL like pattern

[noneck]

Problem

• src/index.ts used the low-level Server API: hand-written JSON-schema ListTools handler, per-case z.object(...).parse, a 7-way switch, and 7 copies of the JSON content envelope (~200 lines of boilerplate).
• Version drift: the Server constructor hardcoded 1.0.0 while package.json says 1.0.1.
• CityRecordNotice declared 32 optional fields that were never field-accessed (results are only JSON.stringify'd).
• buildUrl had a single caller.
• Bug: getNoticesByAgency embedded %25 + encodeURIComponent() inside the SoQL string, which searchParams.set encoded again — Socrata received a literal %25NAME%25 pattern (spaces as %20), not the % wildcard, so agency matches with spaces/special characters silently failed.

Fix

• Migrated to McpServer.registerTool(name, {description, inputSchema: zodShape}, handler) (SDK 1.29.0). Tool names, descriptions, and input schemas are unchanged from a client's perspective. Server construction lives in src/server.ts (buildServer()) so tests can attach an in-memory transport; src/index.ts is now just the stdio entrypoint.
• Version read from package.json via createRequire(import.meta.url) — one source of truth. No version bump.
• CityRecordNotice is now Record<string, string>; the field list is preserved as a doc comment.
• buildUrl inlined into sodaFetch.
• Separate commit: build the SoQL like value unencoded (literal % wildcard per the SoQL like docs, single quotes escaped by doubling) and let URLSearchParams do the single encoding pass.

Test plan (already run)

• Added minimal node:test setup (npm test): builds, then runs test/*.test.mjs.
• test/tools.test.mjs — SDK Client over InMemoryTransport asserts the 7 tool names are unchanged and each has a description + object input schema. Passes.
• test/encoding.test.mjs — mocks fetch, captures the final URL, asserts the decoded $where equals upper(agency_name) like upper('%Parks & Recreation%') and the raw query string contains %25 but not %2525. Written first: fails on the old code (%25Parks%20%26%20Recreation%25), passes after the fix.
• npm run build clean.

🤖 Generated with Claude Code

[noneck]

The double-encoding bug fix (a034f12) has been extracted into its own PR against main (fix/agency-search-double-encoding) so it can merge without waiting on the refactor review. This branch should be rebased once that merges.

[noneck]

Heads up: main has moved — v1.0.2 shipped today (SoQL escaping + NY-timezone fixes from #7 squash-merged, plus a version-from-package.json read and a repository field in package.json required by npm provenance). This branch needs a rebase onto current main before it can proceed; expect overlap in src/index.ts (version wiring) and package.json.

[noneck]

Re-evaluated against current main on 2026-08-09. What still applies from this PR — and, where relevant, which of its claims have since become false — is recorded in #13, so the analysis does not depend on this branch surviving.

This PR remains open and unrebased. Its CI status is from 2026-07-06 against a 2026-07-01 merge base and should not be read as current.