fix(sse): preserve broadcast order and de-race message history

[walcz-de]

What

The SSE broadcast manager runs a pool of N competing delivery workers (NewManager(5) in core/state/pool.go) on a single broadcast channel. Each message is handed to exactly one worker, so with N>1 the delivery order to connected clients depends on goroutine scheduling.

Why it matters

Token-level SSE streams (stream_event chat deltas emitted by the pool's WithStreamCallback) are only meaningful in send order. Under real streaming rates the worker pool visibly scrambles them — in the React UI this surfaces as interleaved/corrupted agent chat output (tokens woven out of order, seemingly "multiple answers braided together").

There are two additional defects in the same file:

• messageHistory.Add() was called once per delivered client inside the per-client loop: with 2 clients each message was recorded twice (duplicated on replay after an EventSource reconnect), and with 0 clients nothing was recorded at all. It also raced unsynchronized across the 5 workers (slice append without a mutex).
• history.Send() wrote to the client channel blocking, so a client with a full buffer could stall the connection handler.

Fix

• Single delivery goroutine consuming a buffered channel: FIFO order is preserved end-to-end, Send() stays non-blocking in practice, and slow clients still drop messages instead of stalling the broadcast (unchanged semantics).
• History records exactly once per message, under a mutex, independent of connected clients.
• History replay snapshots under the lock and delivers non-blocking.

Tests

go test ./core/sse/... -race — adds order-preservation, history-dedup and history-without-clients tests. The order test fails against the previous worker-pool implementation.

🤖 Generated with Claude Code

https://claude.ai/code/session_01KYp5FnAfsdGb6yjLmv8Mge