fix(ctx_shell): detach long foreground runs to a pollable job before the host abort (#1106)

[andig]

Fixes #1106 (and the follow-up in the comment: auto-backgrounded ids that background_action=status reports as not found).

Root cause

The 120s handoff is the MCP host aborting the tool call, not lean-ctx. When it fires it mints a task id in its own namespace; lean-ctx's registry is keyed by shell_* blake3 ids and never sees that id, so background_action=status genuinely can't resolve it. The two namespaces belong to two processes and can't be unified from the daemon side.

should_auto_background only pre-empts at timeout_ms >= 300_000 + cargo test, so any other long command (or a lower timeout_ms) runs foreground and gets stranded by the host abort.

Fix

Add a foreground soft cap below the host abort (default 110s, override LEAN_CTX_SHELL_FG_CAP_MS). Every non-explicitly-backgrounded command now runs as a managed job and is waited on inline up to the cap:

• finishes within the cap → output returned inline, job removed from the registry (unchanged behaviour for the common case);
• outlasts the cap → left running, and its real shell_* id is returned so the caller can always background_action=status/cancel it.

This sidesteps the host abort instead of trying to reconcile the id namespaces.

Notes / caveats

• Relies on the cap being below the host's actual abort; configurable via LEAN_CTX_SHELL_FG_CAP_MS for hosts with a lower cap.
• Cross-process polling of a job started in another session is still out of scope (registry lives in the daemon).

Test

New background_shell tests: fast run returns Finished inline; slow run returns Detached { job_id } with a shell_* id that status resolves. cargo test --lib background_shell — 4 passed, compiled against main.

🤖 Generated with Claude Code

[yvgude]

Cherry-picked to main as ef16916. Thank you for the contribution!