Fix catalog import checkpoint state

[luca-ctx]

Summary

Fixes #16.

This separates catalog completion metadata from incremental import checkpoint metadata for cataloged Codex transcript files.

Problem

When a cataloged transcript changes, catalog_sessions.indexed_status correctly becomes pending, but old indexed_* fields could remain populated. That is misleading because those fields describe the previous file version, not the current cataloged file.

External PR #15 identified the real stale-metadata bug. Its approach of clearing all old indexed_* values on any file change is not sufficient for ctx, because Codex append-only session JSONL refresh was using indexed_file_size_bytes as the tail-import checkpoint. Clearing that value makes append refresh reprocess old events instead of importing only the appended tail.

Data model change

• Keeps indexed_* fields as current-file completion metadata only.
• Adds explicit checkpoint columns on catalog_sessions:
  • last_imported_at_ms
  • last_imported_file_size_bytes
  • last_imported_file_modified_at_ms
  • last_imported_file_sha256
  • last_imported_event_count
• Bumps the history store schema to v14.
• Backfills existing databases from old indexed_* metadata where possible.

Behavior

• Unchanged files preserve completion and checkpoint state.
• Changed files clear stale current-version indexed_* completion metadata.
• Append-only growth preserves a checkpoint only when the previous row was fully indexed to the prior EOF.
• Codex tail refresh reads the explicit checkpoint fields, validates the checkpoint prefix hash when available, and then imports only appended bytes.
• Migrated legacy checkpoints without hashes remain usable for compatibility; the next successful import writes a hash-backed checkpoint.
• Shrink and same-size changed metadata clear the checkpoint so old offsets are not trusted.

Migration notes

The v14 migration adds nullable checkpoint columns and backfills last_imported_* from existing indexed metadata. New successful imports store last_imported_file_sha256 so future tail refreshes can reject rewritten prefixes before using a checkpoint.

Tests

• cargo fmt --all --check
• cargo test -p ctx-history-store
• cargo test -p ctx --test cli search_refresh_auto_tail_imports_appended_codex_session_event
• cargo test -p ctx catalog_import_checkpoint_requires_matching_hash
• cargo test -p ctx-history-capture
• cargo test --workspace