perf: JSONRPCMessage smart union scores all four branches on every inbound message

Description

jsonrpc_message_adapter (src/mcp-types/mcp_types/jsonrpc.py) validates the
JSONRPCRequest | JSONRPCNotification | JSONRPCResponse | JSONRPCError union in
Pydantic's smart-union mode, which attempts/scores every branch for every message.
This adapter is the single decode choke point for every transport:

• server/streamable_http.py (POST body), server/sse.py, server/stdio.py
• client/streamable_http.py (SSE data + JSON responses), client/sse.py, client/stdio.py

The four variants are trivially distinguishable by key presence per JSON-RPC 2.0
(method [+ valid id] → request/notification, error → error, result → response),
so a callable Discriminator + Tags (available at the repo's pydantic floor,
>=2.12.0) selects exactly one branch instead.

Measured impact

scripts/bench_jsonrpc_codec.py (proposed alongside the fix), darwin arm64 / CPython 3.14.2 / pydantic 2.12.5, best-of-5:

For high-throughput agent communication (many small requests/notifications per
second per session) this halves envelope-decode CPU on the hot path of every
transport.

Behavior compatibility

The fix keeps classification parity with today's smart-union outcomes for all
spec-valid messages, verified empirically and pinned by tests:

• {"id": null | 1.5 | true, "method": ...} and absent-id → JSONRPCNotification (current downgrade behavior preserved; note the interaction with #2057 / PR #2075 - if that lands, the discriminator's id-guard becomes the single, clean place to implement the rejection)
• degenerate {method, id, result} → JSONRPCRequest; {id, result, error} → JSONRPCError
• JSONRPCMessage stays a plain union (so isinstance(x, JSONRPCMessage) keeps working); the Annotated/Tag wrapping lives only inside the TypeAdapter
• wire output (dump_json, by_alias=True, exclude_none=True) is byte-identical
• deliberate divergence on spec-invalid hybrids (surfaced by cubic's review on the PR):
  {method, error} hybrids classify as a call instead of JSONRPCError, and
  {method: <non-str>, result} is rejected instead of falling through to
  JSONRPCResponse - smart union picked those via field-count scoring, which let a
  malformed frame masquerade as an error response to a pending request; pinned by tests
• only visible change: ValidationError for unclassifiable input becomes a single
  jsonrpc_message_invalid error with a clear message instead of a 4-branch error dump
  (no in-repo test asserts the old text; error codes and HTTP statuses unchanged)

Additional finding worth recording: with a callable discriminator,
validate_json must materialize the payload to call the discriminator, so the
server's existing two-phase parse (pydantic_core.from_json → validate_python)
is the faster path on large bodies (12.2µs vs 32.1µs on 22KB) - the fix adds a
comment pinning that so it doesn't get "cleaned up" into a regression later.

Proposed fix

PR ready: key-presence callable discriminator on the adapter, zero call-site
changes, new parity/branch tests (100% coverage maintained), standalone
micro-benchmark script.

Disclosure: this issue and the accompanying PR were developed with AI assistance
(Claude Code); all measurements and parity checks were run and verified locally.

References

• Draft PR with the implementation, measurements, and tests: #3135 (kept in draft pending this issue)
• Related: #2057 / #2075 (null-id classification - the discriminator id-guard is the natural locus for that change if it lands)

Commencez dans src/mcp-types/mcp_types/jsonrpc.py, au niveau de jsonrpc_message_adapter, puis examinez les points d’appel du transport mentionnés dans l’issue afin de confirmer qu’aucune modification n’est requise. Exécutez scripts/bench_jsonrpc_codec.py et les tests de parité/branch proposés ; le travail est considéré comme terminé lorsque la validation d’une seule branch, le comportement wire et la parité de classification sont préservés, avec le comportement documenté pour les entrées invalides.