fix: warn before slow CLI commands

[ferueda]

Summary

• print one startup warning for sessions index, sessions data compact, and sessions data repair-orphans when stderr is interactive
• keep redirected and captured invocations silent; stdout, structured output, and exit codes remain unchanged
• safely renew the exact writer owner around potentially long session-replacement transactions
• explain supported recovery when a pre-release library checksum is incompatible

Why

Slow commands need immediate expectation-setting, but a live elapsed counter proved disproportionate for synchronous SQLite work: keeping it responsive required scheduler yields or a worker boundary. This PR intentionally removes that POC. The final UI is one line such as:

Indexing sessions; this may take a couple of minutes.

Live testing also exposed a separate correctness bug: one unusually large Codex session can exceed the 30-second writer lease inside a synchronous replacement transaction. The timer heartbeat cannot run during that transaction, so cleanup could report writer-lease-lost after canonical work committed. Replacement and replacement-failure recording now use the existing fenced leased-transaction primitive, which renews the unchanged exact owner under SQLite's write lock.

Verification

• pnpm check
• 76 test files and 759 tests passed
• focused CLI, indexing, Codex rollout, compact, repair, and writer-coordination suite: 6 files and 98 tests passed
• deterministic lease proofs for a replacement crossing expiry and expired-owner failure recording
• existing stale-owner takeover fencing remains green
• interactive compiled CLI smoke showed the warning before the final result; absent-library compaction created no data
• prior disposable-library replay exercised the large-session lease fix with complete source coverage and a healthy post-run SQLite integrity check