fix: reject compaction waiters when the Codex process exits

[ladydd]

Problem / pressure

/compact waits on a thread/compacted notification. Turn waiters already reject when the Codex process exits (rejectAllPendingTurnCompletions); the comment there says the alternative is a forever hang, then every later prompt is A Codex prompt is already active. Compaction waiters were resolve-only Promises.

If app-server dies after thread/compact/start succeeds, that notification never arrives. The ACP prompt stays occupied. Host Stop/session/cancel does not finish it. The next prompt is already active.

This is not “/compact is broken”: a healthy compact completes. Mid-turn process death already returns -32603 in milliseconds. Compaction is the hole.

This is also not the fork-unsubscribe hang (#37 / Lody#544). Those keep thread/fork children subscribed. This PR does not change fork. Open #30 is stop-during-compact; this is process-exit waiters.

Summary

On connection close/dispose, reject compaction waiters as well as turn waiters. Share the existing connection-terminal error. runCompact still registers the waiter first (so an early thread/compacted is not missed), then Promise.alls start + completion so a close before start returns cannot stall on the in-flight start RPC (vscode-jsonrpc close does not reject responsePromises; dispose does). finally drops this registration only.

Do not fake a successful compact. The error must reach the ACP prompt so the host -32603 recovery path can run.

Visual explanation

sequenceDiagram
  participant Host
  participant Adapter
  participant Codex
  Host->>Adapter: prompt /compact
  Adapter->>Codex: thread/compact/start
  Codex-->>Adapter: start ok
  Note over Codex: process exits
  Note over Adapter: before: wait forever for thread/compacted
  Host->>Adapter: session/prompt
  Adapter-->>Host: already active
  Note over Adapter: after: close rejects compaction waiter, prompt returns -32603

Loading

Before / after

Before	After
/compact + process death: UI stays compacting; Stop then already active	Prompt returns -32603; next prompt is not already active
Healthy /compact	Unchanged
Mid-turn process death	Unchanged (-32603)

Test plan

• vitest run src/__tests__/CodexAppServerClient.test.ts --no-file-parallelism --retry=0 — 7 passed (existing turn-close gap + compact: healthy, early thread/compacted, close after start ack, close while start still in flight, start RPC failure does not leave a waiter, compact after the connection already died).
• Independent check on this patch (Codex 0.153.4, synthetic model, separate bundle): compact-kill returned -32603 in 10ms with no already active; restart + loadSession then history, a normal prompt, and another compact all completed. Lifecycle 8/8 including real vscode-jsonrpc close while start is still pending. Related adapter regression 130 passed; the known /review slash-command timeout is a pre-existing baseline skip, not this change.

Lody still needs a submodule bump after this merge.

Context

Authored with an AI coding agent. Reviewers should challenge: sharing turnCompletionTerminalError as connection-level vs a second compact flag; Promise.all vs awaiting start then completion; whether finally disposing only this waiter is enough when start fails and the process is still alive.

[ladydd]

Rebased onto current main (5f0aab0f, after #39/#40). Same two files; compact waiter tests still 7/7. New head a67f231. Host pin agentclientprotocol#551 retargeted to this SHA.