fix(sandbox): bound ACP sandbox lifetime + self-heal vanished attach

[DIodide]

Why

The per-workspace sandbox unification (#141) creates persistent Daytona boxes with auto_stop_interval but no auto_delete_interval, so Daytona stops → archives → keeps them forever. Two leak sources pile up as archived sandboxes:

1. Leaked session-owned scratch boxes — teardown is missed on a gateway restart/crash, so the box is never deleted.
2. Abandoned workspace boxes — a workspace nobody returns to.

Archived boxes take ~3 minutes to wake (filesystem restored from object storage), so once enough accumulate the whole Daytona account crawls — Claude Code sessions "start indefinitely" and sandboxes won't open. This is exactly what took staging down (82 sandboxes, 60 of them archived). The backlog was cleaned up manually; this PR makes it self-correcting.

What

1. Bound lifetime at the source — set auto_delete_interval on both create paths (provision_agent_sandbox + DaytonaService.create_sandbox):

• Scratch (session-owned) boxes: 1 day — reclaimed before they even archive (default auto-archive is 7 days). They hold nothing durable (deleted on teardown anyway).
• Persistent workspace / code-exec boxes: 14 days — they hold the user's files, so a long grace period.
• The auto-delete clock is the same "continuously stopped" clock that drives auto-archive, so it spans the archived period — archived boxes do get reclaimed (verified against the SDK docstring).
• Both intervals are env-tunable. Non-positive clamps to "disabled" because Daytona reads 0 as delete-immediately-on-stop (a data-loss footgun, not "off"). ephemeral=False on both sites avoids the SDK validator that would force the interval to 0.

2. Self-heal a vanished attach — a box Daytona auto-deletes still looks "owned" to verify_sandbox_owner (which checks the Convex record, not Daytona), so the next session would attach a ghost and hard-error. On DaytonaNotFoundError during an attach, _provision_once now:

• drops the stale Convex link (sandboxes:removeByDaytonaId), then
• for a workspace-unification box, creates a fresh persistent box and relinks it (transparent recovery);
• for an explicit, user-chosen harness sandbox, surfaces the loss (can't fabricate the box the user pointed at) — but the dead link is still cleared so the next session doesn't re-attach it.

This mirrors the existing revive-path heal (which already handles a vanished owned box) and closes the gap on the initial-provision/attach path.

Tests

tests/test_sandbox_lifecycle.py:

• _auto_delete_minutes interval selection (persist vs scratch, scratch < persistent, non-positive → disabled clamp).
• _recover_from_missing_attach decision (workspace box → recover + unlink; explicit harness sandbox → surface but still unlink; id mismatch → treated as workspace).
• _unlink_dead_sandbox calls removeByDaytonaId and swallows ConvexMutationError.

Full fastapi suite: 370 passed, ruff clean.

Notes

• Behavior change for DaytonaService.create_sandbox (Manage Sandboxes / code-exec): a user box untouched for 14 continuous days is now reclaimed by Daytona. Capped at 20/user and env-tunable; raise acp_persistent_sandbox_auto_delete_minutes (or set ≤0 to disable) if a longer-lived box is desired.