Detect stalled and Hono-swallowed SSE writes (CL-7246)

[TheGreatAxios]

Summary

Hono's StreamingApi.write swallows writer errors inside a bare catch {}, so writeSSE can resolve after a failed write or hang forever on a stalled sink. The workbench SSE bridge treated that as a successful delivery and kept the subscriber alive.

Changes

• After every writeSSE, treat an errored underlying writer (desiredSize === null), an aborted or closed stream, or a write that never settles within 10s as a failed delivery.
• Report through reportError, tear down both subscriptions, and close the stream so the client reconnects instead of sitting on a zombie connection.
• Tests cover the swallowed-writer path and the stall timeout (CL-7246).

https://linear.app/abklabs/issue/CL-7246

Fixes CL-7246

[TheGreatAxios]

Fixes CL-7246

[linear-code]

CL-7246

[TheGreatAxios]

Self-review

Verdict: the stall/swallowed-write detection and teardown logic are correct and well-tested against Hono's real StreamingApi (writer.desiredSize, writer.abort() semantics verified against hono's source in node_modules); ready to merge.

• Stalled-write detection via writer.desiredSize === null and the write-timeout race — CONFIRMED, verified against Hono's StreamingApi internals (writer = writable.getWriter()); desiredSize === null is spec-accurate for an errored writer. DISMISSED as a concern.
• Aborting the underlying writer instead of stream.close() on a stall — CONFIRMED, verified this errors the response's pull() read and propagates as a stream error rather than a clean disconnect; already disclosed in the doc comment, and workbench-share-routes.test.ts was updated with a .catch(() => undefined) to account for it. DISMISSED as a concern.
• Overflow path (MAX_QUEUED_DELIVERIES) now also calls dropHungStream instead of a plain closeStream — DISMISSED: queue overflow is itself caused by a task stuck on a hung write, so routing it through the same stall-teardown path is consistent, not a regression.
• report-error-ignore on abortUnderlyingWriter's synchronous-throw catch — CONFIRMED correct usage; bun run scripts/checks/report-error.ts shows 0 new findings.
• Doc comment for bridgeWorkbenchStream had a stray mid-sentence line break ("keeps idle connections alive\nbehind proxies...") — CONFIRMED-FIXED in a follow-up commit.

Tests run (from packages/chat):

• bun test test/workbench-events.test.ts — 25 pass, 0 fail
• bun test test/workbench-share-routes.test.ts — 12 pass, 0 fail
• bun run typecheck — clean
• bunx eslint src/workbench-events.ts test/workbench-events.test.ts — clean