Capture heap visibility and reject incomparable benchmark trials

[bweis]

Benchmarks can otherwise attribute changed heap visibility to an optimization: identical query/settings pairs in our TIN experiments switched from a 16 ms heap scan to a sub-millisecond count path. Record actual visibility-map coverage around each local run and withhold paired ratios when initial coverage differs, evidence is missing, or read-only coverage drifts.

Capture untimed snapshots after setup VACUUM, before driver/warmup, and after the stopped container restarts. Preserve estimated live/dead tuples, maintenance/mutation counters, and table options. Mutation endpoints remain measured outcomes rather than required equal inputs. The report explicitly states that post-restart evidence is not an exact end-of-traffic snapshot or proof of cold-cache conditions.

Based on main after #31, retaining its resource reporting. Requires pg_visibility and fails setup if unavailable. Existing individual artifacts remain readable; historical paired runs need fresh visibility evidence to satisfy the stronger contract.

Validation: 124 Python benchmark tests, 171 source-header checks, and git diff checks pass. Native PostgreSQL18 smoke passed across insert, VACUUM, update, second VACUUM, and restart: coverage0/36 →36/36 →31/40 →40/40 →40/40. Read-only drift was rejected and mutation outcome accepted. This exercises actual capture SQL and comparison logic; it does not claim a full Docker-driver lifecycle run. Evidence is checked into docs/benchmarks/visibility-controls-smoke.json. Ready for review; merge remains gated on all updated-head CI checks passing.