[ppodolsky]

This change adds bounded whole-term wildcard queries and reduces phrase/conjunction and result-projection costs through the existing core readers and scorers. It also adds a reproducible Searchbench HTTP adapter, workload compatibility checks, and measured reports.

• Share prefix/wildcard dictionary expansion with explicit scan, term and posting budgets; preserve logical document IDs after RGB reordering.
• Batch posting intersections, reuse bounded position scratch, and apply conservative phrase score bounds while preserving exhaustive IDs, score bits and counts. New POS5/POS6 position certificates replace legacy position formats: older position indexes require rebuilding.
• Bound the fast-field header directory to 256 checkpoints (3 KiB per reader), move owned envelope strings without allocations, and keep the in-place dispatch experiment opt-in. The corrected 32-client conjunction top-100 comparison improves 4.1% QPS and reduces CPU/request 3.8%; separate ID projection probes improve about 43%. Coverage remains 15 of 826 compatible queries, not the full suite.
• Preserve current main's per-field tokenization, range cursor and broker-budget changes. Fix the existing Python README formatting failure in CI.

Validation: the final merged tree passes all nine full-harness stages (20260923T184501.534018Z-full), including 2,045 native tests (25 normally ignored), native-without-sync, portable core, API docs and real-server broker E2E. GitHub passes the WASM release build and all 40 JavaScript tests on commit 764b7c0b. Earlier byte/score/count audits, diagnostic-feature tests, and HTTP dispatch cancellation/panic/shutdown checks also pass. The benchmark report records exact binaries, memory, rejected experiments and the remaining performance gap. Both benchmark VMs are stopped.

Release: merge before dispatching publish.yml with version_bump=minor, producing 1.9.0 from 1.8.149. The workflow owns package version updates and tags.