Add published-query benchmarks for Stannum and stored-vector GIN

[bweis]

Our existing performance comparisons use a small fixed query set and expression-index GIN. This adds a local entry point around PlanetScale’s pinned ParadeDB Benchmarker fork, replaying the published Wikipedia trace against Stannum and stored-vector GIN with normal planner settings. The upstream driver owns timing and traffic; our wrapper owns setup, correctness, provenance, and reporting.

The integration exposed two issues and retains regressions for both: PostgreSQL position truncation makes it unsuitable as Stannum’s raw-text phrase oracle, and the upstream update picker can select empty heap pages. Per-engine oracles now record cross-engine count differences explicitly, while the picker falls forward/wraps to a live tuple without weakening exactly-one-row accounting. Update runs verify all 906 counts remain unchanged afterward.

Validation: 98 Python harness tests, 6 attribution tests, 8 JavaScript adapter tests, and 7 real-server Go driver tests passed. Native ARM64 PostgreSQL 18.6 runs passed 100k-document mixed counts in both engine orders, 1k-document concurrent updates, and exhaustive same-engine top-10 checks for all 906 forms. Results, limitations, pinned identities, and failure history are recorded in docs/benchmarks/published-trace-results.md.

This establishes the replacement path; it does not delete unique mutation/VACUUM coverage or claim the full benchmark matrix is finished. The guide lists retirement gates for 585 lines of older launchers plus the read-only runner. No extension runtime code changed. No AWS resources were used.