feat(cli): add vfs pack for atomic session transfer artifacts

[factory-ain3sh]

Description

What

Adds vfs pack <session-id>: prepares an inactive run session's delta.db as a single-file transfer artifact. Also adds vfs version --json for capability detection and schema v0.6 with persistent session-handoff metadata.

Why

Session handoff (AC-768) needs a transferable artifact with its own integrity story: a packed database must carry a double-resume guard (generation) and future seed provenance inside the file itself, not in session-directory sidecars that never travel.

How

Pack owns the operation end to end:

1. Takes the exclusive .session.lock; live mounts or owner/joiner procs are rejected with exit code 3.
2. Copies the SQLite family (main + WAL/SHM) to a private staging family; pruning, v0.6 migration, generation bump, checkpoint, and compaction (VACUUM INTO, since Turso lacks in-place VACUUM) happen only on the copy.
3. Prunes generated-artifact globs through the core filesystem API, never via whiteouts, so a pruned base-shadowing path falls back to the base version.
4. Publishes atomically: old family renamed to a deterministic backup, staging renamed into place, metadata verified, rollback on failure. A later pack recovers the backup if the process died between renames. --output artifacts publish via no-replace hard link.

Schema v0.6 adds fs_session_metadata(key, value) (generation, seeded_paths); the v0.5 → v0.6 migration is additive and runs in the same schema transaction.

Stack (merged bottom-up 2026-07-28, vfs main @ 3b2af42):

1. refactor: rename agentfs to vfs across runtime, tooling, and docs #4 rename agentfs to vfs
2. feat(cli): add vfs pack for atomic session transfer artifacts #5 vfs pack <- this PR
3. feat(cli): add vfs seed for portable dirty-base capture #6 vfs seed
4. feat(cli): harden resumable run sessions and add status preflight #7 run hardening + status

Related Issue

Part of AC-768

Reviewer Guide

Diff shape: ~50% cmd/pack.rs (staging + publication state machine), ~20% core (schema/mod.rs, session.rs), ~15% locking (session_lock.rs), ~15% CLI plumbing/docs.
Review depth: Deep on publication ordering; that rename dance is the crash-safety story.
Read order:

1. crates/vfs-core/src/schema/mod.rs — v0.6 migration; confirm it is additive and preserves overlay config
2. crates/vfs-core/src/session.rs — metadata read/write API and its invariants
3. crates/vfs-cli/src/cmd/session_lock.rs — shared-for-lifetime vs exclusive-for-pack lock protocol
4. crates/vfs-cli/src/cmd/pack.rs — staging copy, prune, publish, backup recovery
5. docs/SPEC.md — the failure contract ("a failed pack MUST leave metadata and filesystem rows at pre-pack values")

Open for pushback: pack rejects live sessions outright (exit 3) instead of waiting; the daemon caller owns retry policy.

Risk & Impact

The v0.6 migration runs on every older database pack touches. It is additive (new table only), but a migration bug would be data-corrupting; the SPEC preservation clauses are the contract. Publication has two crash windows between renames; the deterministic-backup recovery path covers both. Prune misconfiguration silently drops delta paths, which is why default prunes are overridable (--no-default-prunes, --prune).

Verification

Behavior verified. Pack on a live session exits 3 with the documented message; interrupted publication recovers the backup on the next pack; --output refuses to overwrite a concurrently created target (verified @ 5ac5526).
Regression coverage. Schema migration tests in vfs-core pin row preservation across v0.5 → v0.6; CLI suites cover the exit-code contract. 1d99611 (post-review) updates test-migrate-consolidation.sh's stale schema-version and user_version expectations for v0.6, caught on the first macOS + ubuntu CI run against main.
Not tested. Concurrent pack from two processes beyond the lock-acquisition test; the exclusive flock makes the race unreachable by construction.
Standard validators. cargo build/test and CLI shell suites clean.