Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@lib/segment/tests/integration/gpu_hnsw_test.rs`:
- Around line 150-153: The test must assert that the GPU HNSW path was actually
used instead of only comparing accuracy: after configuring full_scan_threshold
and performing the search in the GPU HNSW tests, add an explicit assertion that
the GPU/indexed path was exercised (for example, check the runtime indicator
exposed by your codebase such as a Segment/Index method or metric like
is_on_gpu(), used_gpu() or engine.metrics().gpu_search_count > 0); modify the
test cases that use full_scan_threshold (the ones around the shown snippet and
the other ranges called out) to either force the indexed path (adjust
full_scan_threshold appropriately) and then assert the GPU flag/metric is true
so a silent CPU fallback will fail the test.
- Around line 300-309: The assertions currently allow the indexed case to be
worse by up to min_diff; change the comparisons so the indexed hit counts must
exceed the non-indexed ones by min_diff. Replace the check using hits_cpu_idx +
min_diff >= hits_cpu_no_idx with hits_cpu_idx >= hits_cpu_no_idx + min_diff (and
similarly replace hits_gpu_idx + min_diff >= hits_gpu_no_idx with hits_gpu_idx
>= hits_gpu_no_idx + min_diff), and update the assertion messages to reflect the
new ordering (e.g., "Payload index should improve accuracy:
cpu_idx={hits_cpu_idx} < cpu_no_idx={hits_cpu_no_idx} + {min_diff}" adjusted
accordingly).
- Around line 82-97: The test currently compares full ScoredPointOffset lists
for equality; instead extract the returned point identifiers (e.g., map each
ScoredPointOffset to its offset/id) from hnsw_index.search(...) and compare only
IDs to the ground-truth IDs for each query (use a HashSet or sort-and-compare of
IDs) so small score or order differences don’t fail the test; update the closure
in the iterator that calls hnsw_index.search to build a set/vector of result IDs
and compare that to the expected ground_truth IDs (or compute
intersection/recall) rather than comparing ScoredPointOffset structs directly.

---

Nitpick comments:
In `@lib/segment/tests/integration/gpu_hnsw_test.rs`:
- Line 206: This test sets global process state via
set_gpu_groups_count(Some(gpu_groups)); to avoid leaking that change to other
tests, restore the previous state at the end of the test (or use a scope guard)
by calling set_gpu_groups_count(None) (or restoring the original value) — ensure
the cleanup runs even on panic/failure (e.g., using RAII/Drop or a test teardown
helper) so gpu_groups does not remain set for subsequent tests.