Preserve runtime bounds in generic ranked scans

[bweis]

Generic ranked queries with parameterized LIMIT/OFFSET previously lost their top-k bound and ranked all candidates. Bind simple bigint bounds at execution so eligible scans can use existing block-max pruning, rebinding on rescans. NULL, overflow, WITH TIES, and visibility checks retain correct fallback behavior; PostgreSQL still owns negative-bound errors.

A validated 100k Wikipedia A/B/B/A comparison found about 46% lower server latency for broad OR LIMIT 10 and 33% lower for LIMIT 100 OFFSET20. The first implementation regressed a filtered query by 54%; the final change conservatively excludes residual SQL quals, restoring that case to roughly unchanged latency. Existing literal-limit behavior and unsupported expressions/query shapes remain unchanged.

Proof: docs/benchmarks/runtime-ranked-bounds.md and runtime-ranked-results.json include frozen source/image identities, all round medians, plans, and actual visibility-map controls. Both campaigns passed 60/60 exact ordered-score/membership checks. The existing server-time collector now supports prepared SQL cases and alternating measurement order.

Validation: native ARM Docker build passed; 123 Python harness tests and formatting passed. The original implementation passed ARM/x86 PostgreSQL 17/18 tests and upstream Lead checks; final-head CI must pass before merge. Stacked on #32.