Claude Code CLI crashes immediately (exit 1) on every session resume after auto-upgrade to SDK 2.1.111

Title: Claude Code CLI crashes immediately (exit 1) on every session resume after auto-upgrade to SDK 2.1.111

Environment
Claude Desktop app: 1.3036.0
Bundled Claude Code SDK: 2.1.111 (installed at ~/Library/Application Support/Claude/claude-code-vm/2.1.111/claude)
macOS: 26.3.1 (arm64)
Session mode: Local VM (Using Claude VM spawn function for session)
Summary
After the Claude Desktop app auto-upgraded the bundled Claude Code SDK from 2.1.92 → 2.1.111 on restart, every attempt to resume any local VM session crashes the CLI within ~1s of spawn. All crashes surface to the UI as Claude Code process exited with code 1.

The crash is not session-specific — four distinct sessions crashed on resume with identical symptoms. No hooks are configured. ~/.claude/settings.json and settings.local.json are clean.

Timeline (from ~/Library/Logs/Claude/)
Last healthy run on SDK 2.1.92:

2026-04-16 14:53:07 [info] SDK installed successfully: version=2.1.92
2026-04-16 14:53:07 [info] Process spawned: id=55cdde4d-… name=beautiful-zealous-cray command=/usr/local/bin/claude
2026-04-16 14:54:23 [info] Process 55cdde4d-… killed with signal: SIGTERM # normal end-of-turn
App quit + restart triggers upgrade to 2.1.111:

2026-04-16 15:04:05 [info] Stopping VM (isAppQuit: true)...
2026-04-16 15:07:06 [info] SDK installed successfully: version=2.1.111
Crashes immediately follow, across multiple sessions:

2026-04-16 15:07:07 [error] Session local_52ac6f15-… query error: Claude Code process exited with code 1
2026-04-16 15:08:17 [error] Session local_52ac6f15-… query error: Claude Code process exited with code 1
2026-04-16 15:08:38 [error] Session local_0ca0b2f4-… query error: Claude Code process exited with code 1
2026-04-16 15:09:46 [error] Session local_5f9a622f-… query error: Claude Code process exited with code 1
Representative CycleHealth record:

[CycleHealth] Unhealthy cycle: {
session_id: 'local_52ac6f15-97cd-4508-b61f-f2a8c87caef5',
vm_instance_id: '35278aef-c88e-4bbb-9334-a505c5702c34',
model: 'claude-opus-4-6',
cli_session_id: 'fdf3150b-8c4f-4084-ac29-ab11265c9775',
had_first_response: false,
seconds_to_outcome: 6,
is_resume: true,
unhealthy_reason: 'system_error',
error_message: 'Claude Code process exited with code 1',
error_category: 'process_crashed',
transcript_size_bytes: 624067
}
Notes
The transcript JSONL is valid (all 185 lines parse; 171 entries were written by 2.1.92, 14 queue-ops by the app).
cowork_vm_swift.log shows each CLI spawn followed by a SIGTERM ~1s later — i.e. the CLI process exited immediately and the supervisor cleaned up, so no stderr/stack trace is captured. It would be very helpful if the harness could log the CLI's stderr on non-zero exit.
Same /usr/local/bin/claude binary inside the VM is invoked — no user-customized SDK path.
There is no older SDK cached under ~/Library/Application Support/Claude/claude-code-vm/, so no local rollback is possible.
Reproduction
Have any existing local VM session with a non-trivial transcript on SDK 2.1.92.
Let the app upgrade to SDK 2.1.111.
Try to resume the session (send any message).
Observe: CLI exits 1 within ~1s, no first response, is_resume: true.
Impact
All existing local VM sessions are unusable on this machine.