feat: form-mode cost confirmation via MCP elicitations - #367

The hide-tools test added in #334 imports from the v1 monolithic SDK
(@modelcontextprotocol/sdk), which this branch removes. Port it to the
v2 client package and its typed callTool result, matching server.test.ts.

Assert the full sorted tools/list name set, the server identity and
declared capabilities, and one ordinary read-only list_projects call
against a loopback management API stub, so the stable SDK pin flip has
goldens to survive.

Also correct the missing-token comment: the server behavior is
unchanged, only the message the test's own client renders differs
between SDK majors.

Move the catalog from the 2.0.0-beta.3 prerelease to the released
2.0.0 line, and keep the caret range the previous
'@modelcontextprotocol/sdk': ^1.25.2 entry had. The catalog value is
substituted into every published peerDependencies entry, so an exact
pin would hand consumers an unsatisfiable peer whenever they depend on
any other 2.x of @modelcontextprotocol/server.

Also repoint the README samples, which still imported the v1 monolithic
package that this migration removes.

`ExpandRecursively` existed only to widen the intellisense of
`ExtractRequest`/`ExtractNotification`/`ExtractResult`, which this branch
removes because v2's `Server` takes no type parameters. It is still public
through `export * from './types.js'`, so leaving it would publish a type with
nothing left to widen and make its later removal a second breaking release.

Both READMEs' first example imports `@modelcontextprotocol/client`, which no
published manifest declares. Under v1 that resolved for free because
`@modelcontextprotocol/sdk` was the declared peer and npm auto-installs peers;
the v2 split moved the client into its own package, so the example no longer
runs after following the install instructions. Name it in both.

BREAKING CHANGE: `@supabase/mcp-utils` also stops exporting the type
`ExpandRecursively`.

Move the CLI's stdio entry from a one-shot server plus raw
StdioServerTransport onto the SDK's serveStdio, which owns transport
startup, connection-pinned era selection, and teardown. The default
legacy: 'serve' is the dual-era behavior we want, so no option is passed.
CLI parsing, token handling, platform construction, error output, and
exit codes are unchanged.

Export createSupabaseMcpHandler, a thin wrapper over the SDK's
createMcpHandler with legacy: 'reject'. Hosted owns authentication,
ABAC, era dispatch, logging, and request lifetime, so the package
interface carries only the strict modern entry and server construction.

The stdio wire goldens now run as an era matrix. Both eras assert the
same 29-tool set, the same server identity and capabilities, and the
same list_projects content. Modern additionally carries the protocol's
mandatory _meta['io.modelcontextprotocol/serverInfo'] stamp, which
legacy asserts is absent.

Add test:packed-platform-consumer, which packs mcp-server-supabase and
mcp-utils, installs both with plain npm into a throwaway project pinned
to the exact zod version Platform's catalog carries, then asserts the
ESM entry, the CJS entry, the type declarations, and one modern
2026-07-28 call. Registering the account tool group keeps the zod-built
tool surface under test, which an empty catalog would skip.

serveStdio routes transport startup and out-of-band wire errors only
through options.onerror, swallowing them otherwise, so without it a
startup failure was silent. The previous awaited server.connect() surfaced
it through main().catch(console.error).

Restore eager --features validation. Moving server construction into
serveStdio's lazy factory deferred parseFeatureGroups until the first valid
opening message, so an invalid --features value no longer reported and
terminated at startup. Validate the explicitly provided list before serving,
leaving platform-dependent default resolution inside createSupabaseMcpServer.
Covered by a new startup-failure test.

Make the packed-consumer gate prove what it claimed. Its only schema check was
a truthiness test on list_projects, a zero-argument tool whose healthy schema
already carries no properties, so a zod regression emitting property-less
schemas passed. Assert a parameterised witness instead, and drop skipLibCheck
so the packed declarations are really type-checked. Pin the fixture's dev
dependencies exactly and install without lifecycle scripts.

Move the fixture's source out of JS string constants into real files, collapse
the per-check pass-throughs into one runner plus a table, and drop the
esm-entry check that modern-call.mjs already covers by importing and calling
the package ESM entry. The script goes from 429 lines to 240.

Share the msw lifecycle between the two suites that had copied it, assert
status and error code rather than the SDK's exact envelope wording, guard the
integration suite against a stale dist build, and document createSupabaseMcpHandler
with the required per-request mounting pattern.

The modern validation test asserted the SDK's exact envelope message, so an
upstream rewording would fail a test whose contract held. Assert the status,
the JSON-RPC error code, and the structured envelope data instead. The
message is upstream text this package does not own; {key, problem} is
machine-readable and does not churn on rewording.

handler.close() returns a Promise that aborts in-flight exchanges, and the
example discarded it inside an event callback while also making the
createServer callback async, so a cleanup or serving failure surfaced as an
unhandled rejection. Attach a catch to both, and keep the close on res
finishing rather than on the handler resolving, since the latter would cut
streaming responses short.

Address review feedback on #358:

- Comment createSupabaseMcpHandler as modern-protocol only.
- Link the 2026-07-28 release-candidate post beside both copies of the
  protocol revision constant.
- Correct README's handler lifecycle guidance. A shared long-lived handler
  is supported when the platform is deliberately shared; the per-request
  rule applies when platform carries a per-request credential. close()
  ends the handler it is called on, so its timing follows that handler's
  lifetime rather than always being per response.
- Add a CONTRIBUTING testing section covering the suites and the scripts/
  packaging-gate pattern.