fix: preserve native Modal directory snapshots with disjoint mounts

[Filimoa]

Summary

This pull request fixes Modal workspace_persistence="snapshot_directory" unnecessarily falling back to tar whenever any manifest mount cannot be detached, even when that mount is completely outside the directory being snapshotted.

For example, a workspace rooted at /workspace with a Modal cloud-bucket mount at /mnt/remote is a valid, disjoint topology. Modal's cloud-bucket strategy correctly reports that it cannot detach the mount, but the shared fallback check previously ignored the mount target and therefore downgraded Sandbox.snapshot_directory("/workspace") to tar. Because the mount cannot be included in that directory snapshot, its detach capability is irrelevant.

The fallback decision now accepts the resolved directory snapshot root and consults supports_native_snapshot_detach() only for mounts whose normalized POSIX targets overlap that root. Equality, descendants, and ancestors are treated as overlap. Filesystem-wide snapshot callers continue to inspect every mount.

The Modal adapter also preserves the no-copy boundary around mounted storage:

• Detachable overlapping mounts are detached and restored around persistence.
• Modal cloud-bucket mounts equal to, below, or containing the workspace root remain unsupported and are rejected before provider side effects, including trusted-manifest resume.
• Disjoint cloud-bucket mounts remain manifest-owned ephemeral storage and are recreated when a fresh sandbox restores the native workspace image.
• Mount contents, credentials, authority, native snapshot reference formats, restore behavior, and existing tar behavior are unchanged.

Test plan

• Added generic fallback tests covering disjoint, equal, descendant, ancestor, normalized, double-slash, detachable, filesystem-wide, and Windows-flavored typed paths.
• Added Modal provider-fake tests covering native directory-snapshot references for disjoint cloud mounts, overlapping rejection, detach/restore behavior, trusted resume validation, and fresh restore recreation.
• Ran uv run pytest -q tests/sandbox/test_session_utils.py tests/extensions/sandbox/test_modal.py (155 passed).
• Ran .agents/skills/code-change-verification/scripts/run.sh with the Codex sandbox environment and a writable temporary uv cache; formatting, linting, type checking, and the full test suite passed.
• Completed the required independent final review with two clean reviewers on the final patch fingerprint.

Issue number

N/A

Checks

• I've added new tests, if relevant
• I've run .agents/skills/code-change-verification/scripts/run.sh
• I've confirmed all verification steps pass
• If using Codex, I've run /review before submitting this PR