test(vector/hnsw): stabilize incremental_append_recall_matches_fresh_build with best-of-3 (#886)

[mosuka]

Summary

• incremental_append_recall_matches_fresh_build (the perf(vector/hnsw): incremental append onto a small base uses a low-level entry point, degrading concurrent-build quality (seed-then-bulk-load) #872 base>=ef_construction regression gate) occasionally failed on low-core CI runners because the incremental regime runs a structurally different code path from the fresh build (incremental append via promoted_ep vs. one-shot construction), so their variance doesn't cancel the way the rebuild regime (base=1) does.
• Parametrized assert_append_matches_fresh with cold_trials and take the max self-recall@10 across independent seed_then_append builds, applied to the incremental regime only (cold_trials=3); the rebuild regime keeps cold_trials=1 (unchanged — not observed to be flaky).

Why best-of-N, not median or a wider tolerance

• Tolerance widening alone risks also absorbing a real regression (the pre-perf(vector/hnsw): incremental append onto a small base uses a low-level entry point, degrading concurrent-build quality (seed-then-bulk-load) #872 defect measured cold ≈ 0.73).
• Median-of-N could still fail on a single benign-but-unlucky run; max more directly answers "can a good build be reached?" while a true regression still fails every trial (see verification below), so it can't be masked.
• A serial-only fallback would stop exercising the concurrent path perf(vector/hnsw): incremental append onto a small base uses a low-level entry point, degrading concurrent-build quality (seed-then-bulk-load) #872 actually fixed.

Full investigation and design rationale posted to #886 as comments before this PR.

Verification

• RED: disabled the perf(vector/hnsw): incremental append onto a small base uses a low-level entry point, degrading concurrent-build quality (seed-then-bulk-load) #872 promoted_ep fix (temporarily) under RAYON_NUM_THREADS=2 (emulating CI's ubuntu-latest 2-core runner) — all 3 trials produced an identical failing self-recall@10 (~0.83 vs ~0.99 fresh), confirming the pre-perf(vector/hnsw): incremental append onto a small base uses a low-level entry point, degrading concurrent-build quality (seed-then-bulk-load) #872 defect is a deterministic structural funneling through old_ep, not a scheduling fluke, so best-of-3 cannot mask it.
• GREEN: fix restored (writer.rs diff vs. main is empty); incremental_append_recall_matches_fresh_build run 20 consecutive times in release mode under RAYON_NUM_THREADS=2 — all pass (satisfies the issue's acceptance criterion).
• cargo fmt --check clean; cargo clippy -p laurus --all-targets -- -D warnings clean on both 1.95 and 1.97 (CI) toolchains; cargo test -p laurus --lib (1265 passed) and cargo test -p laurus --tests (1467 passed) both green.

Test plan

• cargo test -p laurus --test hnsw_coldstart_recall_test (both regimes) passes
• 20 consecutive runs under RAYON_NUM_THREADS=2 pass
• RED proof: disabling the perf(vector/hnsw): incremental append onto a small base uses a low-level entry point, degrading concurrent-build quality (seed-then-bulk-load) #872 fix fails every trial, confirming no masking
• Full workspace regression (lib + integration + clippy 1.95/1.97) green

Closes #886
Refs #872, #885