[urmzd]

Summary

agentspec session search parsed every candidate transcript one at a time before matching. This change makes full scans 3-8x faster.

• Raw-text prefilter: for literal queries, a transcript is parsed only if its raw file contains every ASCII word (3+ chars) of the query. Regex queries, unreadable files, and non-ASCII words fall through to a full parse, so the check cannot drop a real match.
• Parallel scan: candidates are scanned in rayon chunks and folded in order, so results stay newest first and --limit / --scan still stop early. Prefiltered sessions do not count toward --scan.
• Load by path: SessionMetaIR records source_path (not serialized), and a new SessionAdapter::load_from_meta parses it directly instead of re-walking storage by ID.
• JSON report gains a prefiltered count.

Benchmark (535 sessions, warm cache, release build)

Query	Before	After
no-hit literal	2.75s	0.35s
cargo	2.46s	0.68s
"rate limit" --role user	2.39s	0.91s
--regex 'cargo\s+build'	2.83s	0.53s
--tool-used Bash	2.51s	0.53s

Result sets matched the old binary on all six benchmark queries, except one bug fix: three Gemini files sharing session ID 4e24b246… were previously all loaded from the first file found. They now load from their own files.

Test plan

• cargo fmt --check, cargo clippy --all-targets -D warnings
• cargo test (158 passing, 2 new prefilter tests)
• Old vs new result diff on 6 queries