fix(runtime): deliver StreamStopped with a bounded blocking send

[aheritier]

Summary

Fixes #4136: the TUI got stuck on "Working…" after the stream had actually finished, because finalizeEventChannel emitted the terminal StreamStoppedEvent non-blockingly and silently dropped it whenever the runtime's 128-slot event buffer was full at teardown. Under normal streaming back-pressure (e.g. PersistenceObserver doing a synchronous per-delta SQLite write, or any consumer slower than the model's delta rate) the buffer is routinely full at exactly that moment, so the TUI/supervisor/leantui — which only clear their busy state on receipt of that event — never learned the stream had ended.

Change

• finalizeEventChannel now delivers StreamStopped with a bounded blocking send (~5s deadline, new boundedChannelSink/bounded() helper in event_sink.go, mirroring the existing nonBlocking()). A consumer still draining the channel — even one that already saw ctx cancelled (Esc) but keeps draining to close — reliably receives it. It's dropped only if nothing accepts it within the deadline, i.e. the consumer has genuinely abandoned the channel.
• Deliberately not unbounded: a consumer audit found several call sites that legitimately stop reading before close (a2a, acp, cli runner, LocalRuntime.Run, embeddedchat, server RunSession). An unbounded send there would reintroduce the exact teardown deadlock fix: guard elicitation channel close against in-flight sends #3070 fixed, which decide StreamStopped ordering vs session-end hooks and document teardown trade-offs from #3070 #3074/docs(runtime): document StreamStopped ordering and teardown trade-offs (#3074) #3275 addressed by making this emit non-blocking (and documented delivery as best-effort) in the first place.
• Updated the finalizeEventChannel/StreamStoppedEvent doc comments to describe the new bounded-delivery contract; channel close remains the one guaranteed terminal signal.
• pkg/a2a/adapter.go: the ADK turn-complete event was only ever sent from the StreamStoppedEvent case, so a dropped event left the turn hanging. Added a channel-close fallback so it still completes if that event was never delivered.

Tests

• New pkg/runtime/stream_stopped_delivery_test.go: a deterministic regression test reproducing the TUI shows "working" even after stream has ended for Qwen 3.8 (served locally via vLLM) #4136 race (slow EventObserver + ≥250 streamed deltas) asserting exactly one StreamStopped, last before close; plus a "slow-but-alive consumer eventually receives it" case. Passes -count=20 with zero flakes.
• Reworked TestLocalRuntime_FinalizeEventChannelDoesNotDeadlockWhenBufferFullAndConsumerGone → TestLocalRuntime_FinalizeEventChannelDropsStreamStoppedAfterBoundedTimeout: injects a short timeout, asserts the send waits out the deadline (proving bounded, not instant-drop) and still returns/drops for a genuinely abandoned consumer.
• TestLocalRuntime_FinalizeEventChannelStreamStoppedIsLastBeforeClose / ...EmitsStreamStoppedOnce pass unchanged.

go build, go test ./..., golangci-lint run (v2.13.1) all clean. (pkg/rag/treesitter fails in this sandbox only — pre-existing, missing gcc for CGO, unrelated to this change.)

Independent review (sub-agent) returned only low-severity, non-blocking nits (comment verbosity, a defensive suggestion for an unreachable wait<=0 case, a minor doc-wording ambiguity) — none required changes.