build(sandbox): pin Zed upstream catch-up (764 commits) + silent-agent watchdog

[lukemarsden]

Pins ZED_COMMIT to the Zed-side upstream catch-up + silent-agent watchdog, and adds the incident write-up.

Paired Zed PR: (linked below once opened — opening this one first per the ordering rule in CLAUDE.md, so CI can never end up pinned at a stale commit if the Zed PR merges first.)

What's in the pinned Zed commit

1. 764-commit upstream catch-up. The fork had drifted 41 days behind (last fence e45e42af6e, 2026-06-18). It is now level with upstream/main at b9256fa8f0. Merged in four dated rounds rather than one jump, so every conflict is attributable to a bounded upstream window:

Round	Range	Commits	Conflicts
1	e45e42af6e..4b7369481d	254	6
2	4b7369481d..4a3e0af532	212	3
3	4a3e0af532..d23aaeebea	226	1
4	d23aaeebea..b9256fa8f0	72	2

2. A fix for the 2026-07-29 production wedge (spt_01kyq6qd8h4wqq2rfasc4b01e9): claude-agent-acp accepted a session/prompt and went permanently silent — no Stopped, no error, so no message_completed and no chat_response_error. The Helix interaction sat in state=waiting indefinitely with nothing surfaced in either UI. Only a container restart recovered it.

Nothing caught it: the connection.prompt() await is unbounded, and the existing wedge recovery keys on an ede_diagnostic error string that a silent wedge never produces.

The fix is a time-to-first-event watchdog, deliberately not a turn timeout — a blanket timeout false-positives on long tool calls, but "zero events since dispatch" is unambiguous. Once the agent emits anything it disarms for the rest of the turn. On expiry it raises a wedge error that feeds the existing force-close → reload → retry machinery rather than a new mechanism. It lives entirely in the Helix-owned crate, so the upstream diff stays at zero.

Validation

• cargo check --package zed --features external_websocket_sync — clean
• cargo test -p acp_thread — 127 passed, 1 ignored (documented Critical Fix fix bug where sdxl finetune is running and base sdxl inference request hangs forever #8 divergence)
• cargo test -p external_websocket_sync — 55 passed
• Drift sweep — 28/28 after every round
• Dockerized E2E on the final tree — PASSED for both agent rounds (17/17 phases each, store validation passed, response-entries isolation clean across 34 interactions in 10 sessions)

Two pre-existing breakages repaired (neither caused by this merge)

• A raw, unresolved conflict block had been committed into Zed's .gitignore by an earlier merge — markers and all.
• cargo test -p acp_thread had been red since 2026-06-18. An upstream compaction test asserts a displaced turn still delivers its PromptResponse, which is mutually exclusive with Critical Fix fix bug where sdxl finetune is running and base sdxl inference request hangs forever #8 (cancel() drops send_task so a buggy ACP agent can't wedge the next turn). Confirmed failing identically on the pre-merge commit via a worktree, so not a regression. Marked #[ignore] with full rationale; Helix keeps Fix fix bug where sdxl finetune is running and base sdxl inference request hangs forever #8. It hid for six weeks because the rebase checklist only ever ran a single filtered test — the checklist now requires the full suites.

Known gaps (documented, not fixed here)

• The watchdog covers "zero events after dispatch". It does not cover "streamed some output, then stalled before completing" — that shape disarms it. Seen once in the E2E claude round; needs a separate idle-since-last-event budget.
• threadIsWedged() in session_handlers.go only classifies State == error, so a hung waiting turn is still invisible to it.
• Zed's e2e-test/run_e2e.sh queries the wrong npm scope for the agent version, so its attribution log line is always unknown.

🤖 Generated with Claude Code

[cursor]

Bugbot is not enabled for your account, so this pull request was not reviewed.

Enable Bugbot in the Cursor dashboard to get automatic reviews on future PRs.

[lukemarsden]

Paired Zed PR: helixml/zed#73

Merge order (per CLAUDE.md): Zed PR #73 first, then this one.