#22480: fix: memory leak, silent WS failures, and connection error handling

## Summary - **Fix memory leak in `seqByRun` map** (`src/infra/agent-events.ts`): The `seqByRun` map grew unboundedly with every agent run but was never cleaned up. Now `clearAgentRunContext()` and `resetAgentRunContextForTest()` also clean up the corresponding `seqByRun` entries, preventing memory exhaustion in long-running gateway instances. - **Log WebSocket send failures** (`src/gateway/server/ws-connection.ts`): The `send()` function was silently swallowing all errors with an empty catch block. Now it logs a warning with connection ID and error details, and records `"send-error"` as the close cause for diagnostics. - **Handle all socket errors, not just the first** (`src/gateway/server/ws-connection.ts`): Changed `socket.once("error")` to `socket.on("error")` so subsequent socket errors are also caught and trigger connection cleanup. The `close()` function is already idempotent (guards with `if (closed) return`), so repeated calls are safe. - **Fix voice transcript deduplication cleanup** (`src/gateway/server-node-events.ts`): The cleanup loop had an early `break` that stopped pruning expired entries as soon as the map dropped below `MAX_RECENT_VOICE_TRANSCRIPTS`, leaving stale entries behind. Removed the early break so all expired entries are pruned before falling through to the eviction-by-insertion-order loop. ## Test plan - [x] `vitest run src/infra/agent-events` — 3/3 tests pass - [x] `vitest run src/gateway/gateway-misc.test.ts src/gateway/client.test.ts src/gateway/chat-sanitize.test.ts` — 31/31 tests pass - [x] `tsc --noEmit` — no new type errors - [ ] Verify gateway stability under sustained load (memory usage should plateau instead of growing) - [ ] Verify WebSocket send failures now appear in logs with `send failed conn=` prefix 🤖 Generated with [Claude Code](https://claude.com/claude-code) <!-- greptile_comment --> <h3>Greptile Summary</h3> Fixes three critical infrastructure issues in the gateway WebSocket handling and agent event management: - **Memory leak in `seqByRun` map**: The map tracked sequence numbers for all agent runs but was never cleaned up. Now properly deleted in both `clearAgentRunContext()` and `resetAgentRunContextForTest()`, preventing unbounded memory growth in long-running gateway instances. - **Silent WebSocket send failures**: Previously swallowed all errors with an empty catch block, making debugging impossible. Now logs warnings with connection ID and error details, and records `"send-error"` as close cause for diagnostics. - **Incomplete socket error handling**: Changed from `socket.once("error")` to `socket.on("error")` to handle all socket errors, not just the first one. The existing `close()` function is already idempotent (guards with `if (closed) return`), so repeated calls are safe. - **Voice transcript deduplication cleanup bug**: The cleanup loop had an early `break` that stopped pruning as soon as the map size dropped below `MAX_RECENT_VOICE_TRANSCRIPTS`, leaving expired entries behind. Removing the early break ensures all expired entries are pruned before falling through to the eviction-by-insertion-order loop. All changes are well-tested and follow the repository's coding conventions. The fixes address real production issues that would cause memory leaks and silent failures. <h3>Confidence Score: 5/5</h3> - Safe to merge with no risk - fixes critical production bugs - All changes are simple, well-reasoned bug fixes with clear benefits. The memory leak fix properly mirrors the cleanup pattern used for `runContextById`. The WebSocket error handling improvements add observability without changing behavior. The voice transcript cleanup fix removes a logic error. Tests pass and the changes are minimal and focused. - No files require special attention _{Last reviewed commit: 64a049c} <!-- greptile_other_comments_section --> _{(2/5) Greptile learns from your feedback when you react with thumbs up/down!} <!-- /greptile_comment -->