Skill: engram-live-verify

Unit + integration tests prove wiring; live verification proves behavior. Run the actual server binary against an actual store and call the actual tools. This is non-negotiable for scan/retrieval/MCP changes.

The driver

cargo build --release -p engram-mcp   # (add --features pgvector,fastembed for PG/vector work)
python3 scripts/dev/mcp_driver.py \
  --storage ~/.engram/<store> --project <project> [--no-vector] [--wait 12] \
  call <tool> '<json-args>' \
  call <tool> '<json-args>'

Timed responses + captured stderr. Stores on this machine: agentzero (shared, mixed repos + session memories — do not wipe), mem-alpha-self (this repo only), springboot-demo, dreamhouse-lwc (single-repo demos; wipe-able after killing the BFF on :3001).

The scars this checklist is made of

• --wait 12+ when vector is on — FastEmbed model load takes seconds; requests sent too early queue behind boot and pollute timings.
• stderr is where boot lives. knowledge_store=false in a bootstrap warning means silent capability degradation — a store that "boots" can be missing its knowledge handle. Always read the STDERR section.
• Silent-empty is the default failure mode. A query returning []/0 may be correct OR may be: wrong scope semantics (IS NULL OR = is NOT scope-allows — NULL is not a wildcard), a camelCase JSON path probe ($.documentId, never $.document_id), a swallowed unwrap_or_default, or a schema drift. Cross-check counts with sqlite3 CLI before believing empty.
• The vec0 vector table is a virtual table — plain CLI reads of vectors_vector_chunks* shadows mislead (inserts "vanish"); the vectors_rowids shadow holds the id map. The tool's own embedded/reused accounting is authoritative for reindex.
• Wiping a store: kill the BFF (lsof -ti :3001 | xargs -r kill) first — an open handle blocks rm -rf silently. rm -rf failures print nothing when backgrounded; verify with [ -d path ].
• Heredoc-python drivers get their logs eaten by the shell harness — use script files (scripts/dev/mcp_driver.py exists for this reason).
• Postgres work: docker compose -f docs/how-to-pg/docker-compose.yaml up -d; TRUNCATE tables between tests; the docker-gated recipe tests run with cargo test -p engram-backend-pgvector -- --ignored --test-threads=1.

What "verified" means

1. The tool call returned (not timed out) with the expected shape.
2. Counts changed where they should (scan → entities/rels; reindex → embedded/reused split; embed → the vectors accounting).
3. The invariant held: resolved-signal counts stay stable when only noise is removed (e.g. resolved calls before/after a filter change).
4. For perf claims: before AND after numbers from the same store, same call, same cap — single-sided timings are not evidence.

Record the verification (commands + numbers) in the commit message.