v0.23.0

Breaking Changes

• the opt-in LLM summary enrichment subsystem is removed entirely: ccrecall backfill llm-summaries, the ccrecall-llm-summaries console script, and its six llm_summary_* config keys are gone, and display_title/summary_preview no longer appear in ccrecall --json search/search-messages output. Unknown keys left over in config.json are ignored, so stale config entries are harmless (#134)

Bug Fixes

• surface previously-silent parsing, database, and hook failures to the per-process log files instead of failing silently (#154)
• fix a memory spike in sync-current that could freeze the machine (7.5 GB observed) by capping sync-path embedding at 4096 tokens, dropping the worst-case peak to ~4 GB (#166)
• fix exclude_projects failing to match hyphenated project names on the lossy project-name fallback path (#135)
• fix the next schema migration bricking every database that has ever produced embeddings, by loading the sqlite-vec extension before every migration instead of only the first one; also fix an infinite loop on a corrupted transcript, and a non-atomic handoff-file write that could lose the handoff on a fresh machine (#132)
• warn when a database's schema version is ahead of what the running code expects, instead of silently treating it as fully migrated (#167)

Refactoring

• drop the redundant --n long-form alias from search/search-messages/tail (keep -n); rename recent's --n to --limit; add --verbose to search-messages and --db to all backfill subcommands (#150)