feat(transport): adopt v3 WebSocket subprotocol

[infinia-yzl]

Summary

Flips the WebSocket subprotocol identifier from v2.bsatn.spacetimedb to v3.bsatn.spacetimedb to opt into the v3 transport introduced in clockworklabs/SpacetimeDB 2.2.0 (PR clockworklabs/SpacetimeDB#4761).

Why this works as a one-line change

v3 is not a new message schema. It reuses the v2 ClientMessage / ServerMessage values as-is. The only on-the-wire change is that a single WebSocket payload may now contain one OR more BSATN-encoded v2 messages concatenated back to back. Senders may coalesce as they please; receivers must decode sequentially until end-of-buffer.

The existing BSATNDeserializer.process_bytes_and_extract_messages already loops decoding messages from the incoming buffer until empty, so the receive-side semantics work for free. The send side keeps sending one logical message per frame, which is still valid v3 — a future optimization is to coalesce client-side sends, deliberately not part of this PR.

Verification

Server handshake responds with 101 Switching Protocols + sec-websocket-protocol: v3.bsatn.spacetimedb:

curl -s -i -N \
  -H "Connection: Upgrade" -H "Upgrade: websocket" \
  -H "Sec-WebSocket-Version: 13" \
  -H "Sec-WebSocket-Key: dGhlIHNhbXBsZSBub25jZQ==" \
  -H "Sec-WebSocket-Protocol: v3.bsatn.spacetimedb" \
  -H "Authorization: Bearer <token>" \
  http://localhost:3333/v1/database/<name>/subscribe

End-to-end verified in a downstream project via integration tests against a live SpacetimeDB 2.2.0 dev DB:

• anonymous connect + subscribe + iter() round-trip ✓
• overlapping-subscription audit (subscribe narrow, then subscribe wide; assert on_insert fires exactly once for the new row, zero additional times for the cached row — analogous to the Unreal SDK fix in fix(unreal): suppress spurious OnInsert on overlapping subscription refcount bump clockworklabs/SpacetimeDB#4903) ✓

Test plan

• Existing test-server-module flows continue to pass against a 2.2 server
• No regressions in the procedure / reducer / subscription codepaths (none expected — schema is unchanged)
• Hand-test connect-and-subscribe in any consuming project

Notes for reviewers

• Not adding client-side outbound batching in this PR. Server batches naturally during high-throughput periods; client-side batching is a separate optimization that's worth measuring before adding complexity.
• The version: String = "v1" on line 10 is the HTTP path prefix (/v1/database/...), unrelated to the WebSocket subprotocol version — leaving untouched.

🤖 Generated with Claude Code