feat: add owner-coordinated database compaction

[vinceblock99]

Repeated writes can leave reclaimable space in the combined atomic.redb. Add atomic compact [--repository PATH] [--json] to let users explicitly reclaim that space through the existing database owner. The command reports the absolute database path, before/after file sizes, and reclaimed bytes.

This follows the database-size investigation: on a copy of a 66-file repository, explicit compaction reduced the file from 19,664,896 to 14,307,328 bytes, reclaiming 5,357,568 bytes (27.2%) while preserving its data. redb already reuses freed pages and can shrink on ordinary commits; explicit compaction can reclaim additional space by reorganizing pages. This PR addresses that measured opportunity without changing the graph/index representation or deleting change objects.

Compaction requires exclusive access, so it is an explicit owner operation rather than maintenance on every command or shutdown:

• The owner blocks new store leases, drains active requests, and acquires redb's exclusive file lock. Existing reader/writer handles cause a bounded wait using ATOMIC_DB_LOCK_WAIT_MS, then a clear error if still busy. Once started, compaction runs to completion.
• The database opener lives in the owner's private maintenance module (pub(super)), avoiding a new public Repository opening API. The CLI sends an IPC request; the owner remains available afterward.
• Sandbox invocations resolve to their canonical database. An additive capability check tells users to restart an older owner before retrying.
• Size reporting includes redb's allocator-metadata write on close, so it reflects the final file size. A repeat compaction may succeed with zero bytes reclaimed.

Graph, views, provenance, .change files, and unrecorded working files are preserved. This command only operates on an existing atomic.redb: it does not create tables/databases, migrate legacy layouts, delete persistent savepoints, or run automatically. Large repositories should be compacted during a quiet period because other database operations may exhaust their lock-wait budgets while compaction runs.

This Draft PR is stacked on #230 and targets feat/atomic-redb, so its diff contains only compaction and its tests/documentation. Retarget to dev after #230 lands. This PR does not modify atomic-storage.

Validation on macOS:

• cargo test --workspace --locked --offline: 9,125 passed, 0 failed, 243 ignored, including all 28 owner integration tests.
• Seven new unit tests and two new process integration tests cover actual reclamation, retained data, missing databases, busy reader/writer handles, future schema rejection, persistent savepoint preservation, owner capability compatibility, lease draining, sandbox resolution, and timeout/retry.
• The 66-file CLI test preserved all working-file and change-file SHA256 hashes, history, views, and diff output. Restoring a modified recorded file succeeded; a second compact reclaimed zero bytes. The original repository was unchanged.
• cargo clippy --workspace --locked --offline -- -D warnings, cargo fmt --all -- --check, and the CI semantic diff harness (23/23 assertions) passed. Linux and Windows verification is left to CI.