fix: repair transcript skill evidence and search

[jmagar]

Summary

• recognize modern package-qualified Claude skill command envelopes in live ingest and historical backfill
• version transcript extractor semantics and safely replay stale sources, removing stale skill/MCP/hook projections
• preserve Codex user/assistant speaker roles and require user-authored evidence for user_correction_after_skill
• restore MCP schema/CLI parity for specialized skill/MCP/hook incident filters
• move search_sessions onto a transcript-only ai_logs_fts projection instead of the global log FTS corpus
• keep the AI FTS exact with transcript insert/delete triggers and maintenance support

Production-scale diagnosis

• live corpus: about 55.7M log rows / about 141 GB SQLite+FTS
• seven-day AI transcript slice: about 188k Claude + about 163k Codex events
• old narrow search_sessions calls timed out around 25-30s while ordinary FTS searches returned in about 0.4-2.1s
• root cause was global FTS candidate walking before selective AI filters

Regression coverage

• modern Claude command extraction and unqualified-command rejection
• live scanner insertion and historical backfill
• extractor-revision checkpoint replay
• source replay clears skill/MCP/hook projections
• assistant acknowledgments cannot count as user corrections
• MCP specialized-filter schema plus CLI completion parity
• transcript-only FTS insert/delete behavior and all existing search_ai_sessions semantic regressions
• migration recovery fixtures remain convergent

Verification

• CORTEX_FULL_PRE_PUSH=1 cargo xtask pre-push passed on the final tree
• cargo fmt -- --check passed
• git diff --check passed
• cargo xtask check-version-sync passed (18 carriers at 3.16.2)
• targeted regression matrix passed

Beads: unraid-mcp-oyf7