[bench] Lex-sorted-string ORDER BY fast-path regression bench

[aaj3f]

Context

Follow-up bench from PR #1228 review (item 3 in coverage gaps). The 200× ORDER BY ?label LIMIT speedup (memory: mem:fact-01kkq57f5xpx815scy9fsb6jzr) depends on the lex_sorted_string_ids flag being preserved through bulk import + the first post-import write. A regression here would silently regress production ORDER BY queries on freshly imported datasets — the failure mode is "code still runs, query just gets slow," which is exactly the class the gate exists to catch.

Scope

Add fluree-db-api/benches/query_orderby_strings.rs (alternatively rolled into import_bulk.rs as a follow-up scenario, but a separate file is cleaner). Specifically:

• Setup: bulk-import a dataset with string-typed predicate values that exercise the lex-sorted-strings fast path. Use gen::people or gen::bsbm (whichever's distribution stresses it best) — pick deterministically so the baseline is stable.
• Two scenarios:
  • orderby_freshly_imported: ORDER BY ?label LIMIT N immediately after a bulk import (the lex_sorted_string_ids flag should still be set).
  • orderby_after_first_write: same query, but after a single post-import write txn (which clears the flag per the invariant doc). The bench documents the latency cliff so a regression that fails to clear the flag AND a regression that clears it prematurely are both visible.
• Throughput metric: ns/query or rows/sec (criterion default vs. Throughput::Elements).
• Use the chassis pattern. bench_runtime(), current_profile(), current_scale(), next_ledger_alias().
• Register a budget in regression-budget.json. Tighten the budget on the post-import scenario specifically — that's the one most prone to silent regression.

Acceptance

• Bench compiles and runs --test green at tiny scale.
• regression-budget.json has an entry.
• Both scenarios show clearly-different latencies in a manual cargo bench run (verifies the bench is actually exercising the flag-driven path, not just measuring noise).
• BENCHMARKING.md's "Current benches" table gets a row.
• Bench-gate CI job picks it up.

References

• PR Workspace-Wide Benchmarking Foundation #1228 review: coverage gap #3.
• Memory: mem:fact-01kkq57f5xpx815scy9fsb6jzr (lex-sorted-strings ORDER BY 200× speedup).
• Adjacent bench: import_bulk.rs covers the import side; this bench covers the consequence on subsequent reads.

Out of scope

The deeper correctness invariant (when should the flag be set/cleared) is the subject of existing tests. This bench is purely about catching perf regressions that flip a known fast-path off.