fix(core-internal): silence unknown-format warnings in v2 validators

[senorMk]

Refs #2855 (v2).

Servers that publish unknown format annotations cause the v2 default Ajv validators to print warnings while compiling tool schemas. A three-field schema emits six warnings per compilation: two for each subschema path. The shared default-engine factory now filters these warnings for every supported dialect. Known formats still validate, other diagnostics still reach the console, and caller-supplied engines keep their own logging.

#2898 covers the corresponding v1.x fix. This targets main and complements that work. #2601 also targets v2, but registers google-duration specifically; this change handles arbitrary unknown formats without adding vendor-specific validation. Both release lines need the fix, so this PR references rather than auto-closes #2855.

Adapted from sean-park-funda's v2 patch, 36da288, with co-author credit. The logger stays internal, and regression coverage now exercises the default, 2020-12, 2019-09, draft-07, and draft-06 dispatch paths.

Validation:

• The five new silence tests fail on unmodified main, each observing 12 warnings across two distinct schema compilations, and pass with the fix.
• All 1,541 core-internal tests pass, including 16 new cases covering zero warnings, duplicate diagnostics from custom engines, schema validation, known formats, and forwarding other logger messages.
• core-internal typecheck and ESLint on both changed TypeScript files pass; changed files are Prettier-clean.

The normal pnpm install could not access the registry in this environment. Checks used the repository's lockfile-pinned dependencies downloaded separately and verified against their integrity hashes. The full monorepo checks were not run.

[changeset-bot]

🦋 Changeset detected

Latest commit: 0210f84

The changes in this PR will be included in the next version bump.

This PR includes changesets to release 6 packages

Name	Type
@modelcontextprotocol/core-internal	Patch
@modelcontextprotocol/client	Patch
@modelcontextprotocol/server	Patch
@modelcontextprotocol/codemod	Patch
@modelcontextprotocol/core	Patch
@modelcontextprotocol/server-legacy	Patch

Not sure what this means? Click here to learn what changesets are.

Click here if you're a maintainer who wants to add another changeset to this PR

[pkg-pr-new]

Open in StackBlitz

@modelcontextprotocol/client

npm i https://pkg.pr.new/@modelcontextprotocol/client@3007

@modelcontextprotocol/codemod

npm i https://pkg.pr.new/@modelcontextprotocol/codemod@3007

@modelcontextprotocol/core

npm i https://pkg.pr.new/@modelcontextprotocol/core@3007

@modelcontextprotocol/server

npm i https://pkg.pr.new/@modelcontextprotocol/server@3007

@modelcontextprotocol/server-legacy

npm i https://pkg.pr.new/@modelcontextprotocol/server-legacy@3007

@modelcontextprotocol/express

npm i https://pkg.pr.new/@modelcontextprotocol/express@3007

@modelcontextprotocol/fastify

npm i https://pkg.pr.new/@modelcontextprotocol/fastify@3007

@modelcontextprotocol/hono

npm i https://pkg.pr.new/@modelcontextprotocol/hono@3007

@modelcontextprotocol/node

npm i https://pkg.pr.new/@modelcontextprotocol/node@3007

commit: 0210f84

[chrikrah]

@senorMk, the spam is real, and most of it lands where nobody can fix it: client.ts:1846 compiles every third-party tool's outputSchema through the default engine, so a client author sees the vendor's google-duration and can't change it. non-blocking: I'd still warn once per format name per engine, not never. Silence also removes the only typo signal a server author gets through fromJsonSchema or an elicitation requestedSchema (server.ts:1224), and Ajv then accepts any string. A caller-supplied engine keeps its warnings under this PR, but the author who typed date_time never configured one. The cost of my version is one unfixable line per vendor format on each client.

Put the Set inside createDefaultAjvInstance. A module-level one would also mute a second client meeting the same name, and it would leak across the five dialect tests. Probe: one default AjvJsonSchemaValidator, two tool schemas with three google-duration fields each, then format: "date_time" validating "not a date". It's a scratch test outside the diff:

$ npx vitest run test/validators/zzprobe.test.ts
== base b022522
PROBE warnings=14 typo-valid=true formats=["google-duration","date_time"]
== PR 0210f84
PROBE warnings=0 typo-valid=true formats=[]
== once-per-name, Set per engine
PROBE warnings=2 typo-valid=true formats=["google-duration","date_time"]

The new tests discriminate, and with the per-engine Set each dialect's suppress test sees exactly two warnings:

$ npx vitest run test/validators          # at 0210f84
 Test Files  3 passed (3)
      Tests  125 passed (125)
$ git apply -R src.diff && npx vitest run test/validators
 Test Files  1 failed | 2 passed (3)
      Tests  5 failed | 120 passed (125)
$ npx vitest run test/validators/ajvUnknownFormat.test.ts --reporter=json   # Set per engine; failures
["unknown formats ('default')"] ['Number of calls: 2']
["unknown formats ('2020-12')"] ['Number of calls: 2']
["unknown formats ('2019-09')"] ['Number of calls: 2']
["unknown formats ('draft-07')"] ['Number of calls: 2']
["unknown formats ('draft-06')"] ['Number of calls: 2']

@felixweinberger, this is a policy call more than a code one. I would keep one line per format name for the typo signal; do you want that, or should unknown formats stay silent? The answer should bind the open v1.x fix in #2898 too.