fix: recover forwarded Claude skills and scope session search

[jmagar]

Why

Production verification after #247 exposed two gaps that CI did not model:

1. Forwarded Claude history is privacy-scrubbed. ai_transcript_path is a SHA-256 locator and raw is empty, so historical skill backfill cannot reopen the original JSONL. Modern package-qualified command envelopes still survive in normalized user message content, but the backfill previously ignored that evidence.
2. The transcript-only FTS from migration 60 still evaluated common terms globally before relational tool / since filters. On production, search_sessions(query="error", tool="claude", since=2026-09-11, limit=5) still exceeded the gateway timeout at ~59s.

Fix

Forwarded Claude command recovery

• Recover package-qualified <command-message> / <command-name> evidence directly from persisted normalized Claude user messages.
• Require persisted metadata_json.event_kind == "user" before trusting forwarded command content.
• Preserve the existing source-file fallback for structured attributionSkill / attributionPlugin evidence.
• Preserve source_unavailable accounting when neither normalized command evidence nor the original source is recoverable.
• Add a production-shaped regression with scrubbed metadata, SHA-256 transcript locator, empty raw payload, and vibin:review-pr command content.

Scoped session search

• Add migration 61.
• Rebuild ai_logs_fts with provider scope (message, ai_tool).
• Add partial AI-only idx_logs_ai_timestamp_tool(timestamp, ai_tool) WHERE ai_tool IS NOT NULL.
• Scope the FTS expression to message and, for safe provider identifiers, ai_tool.
• For since, derive an exact-safe minimum rowid from the AI-only timestamp index and push that bound into FTS before common-term expansion.
• Retain the existing relational l.ai_tool = ? and timestamp filters as correctness guards.
• Do not interpolate unsafe provider values into FTS; they fall back to the parameterized relational equality filter.

The rowid floor is derived from actual rows satisfying the event-time bound, not inferred from ingestion order, so late-arriving historical transcripts remain discoverable.

Production evidence before this PR

• Cortex fix: repair transcript skill evidence and search #247 deployed cleanly as 03ba04df.
• Migration 60 completed on the ~160 GB production DB in 146.0s; service returned healthy with zero restarts.
• Modern Claude skill_events since 2026-09-01: 0.
• Production forwarded Claude rows preserve command envelopes in message but have scrubbed source/raw data.
• Representative scoped search_sessions call: ~58.996s timeout.
• Live query plan showed FTS driving globally, followed by rowid lookups and relational filtering.

Validation

TDD:

• New forwarded/scrubbed Claude regression failed first with inserted=0, source_unavailable=1, then passed after the recovery path.
• New migration 61 and scoped-session-search regressions failed before implementation, then passed.

Targeted matrix:

• all skill_backfill tests
• all search_ai_sessions tests
• migration 51 damaged-schema recovery
• schema-43 historical upgrade fixture
• migration 61 provider-scoped FTS/index regression
• schema/docs parity
• unsafe tool-scope interpolation guard

Full isolated gate:

CORTEX_FULL_PRE_PUSH=1 cargo xtask pre-push

Main library suite:

3030 passed
0 failed
3 ignored

All integration/workflow suites also passed.

Deployment verification after merge

After hosted CI is green and this PR is merged:

1. build/deploy the exact merge SHA to TOOTIE with the current image rollback preserved;
2. observe migration 61 to healthy completion;
3. run the production skill backfill and verify modern Claude claude_skill_command events become nonzero;
4. rerun the exact pre-deploy search_sessions query and record before/after latency + result correctness;
5. update U8-929 with landed production evidence.