fix(query-engine): run top-k CMSHeap on FastPath

[milindsrivastava1997]

Closes #811

Runs the query engine top-k CMSHeap directly on CMSHeap<Vector2D, FastPath>.

• preserves count and rounded integer sum updates
• removes legacy portable deserialization compatibility
• accounts for i32 counters in memory estimates
• adds a fixed-stream FastPath top-k regression test

Tests:

• cargo test -p query_engine_rust fastpath_fixed_stream_matches_expected_estimates_and_topk --lib
• cargo test -p query_engine_rust cms_with_heap --lib
• cargo test -p query_engine_rust --test e2e_precompute_equivalence

[milindsrivastava1997]

Automated code review (Claude Code). Main theme: the f64 → i32 counter switch changes SUM top-k semantics (fractional values dropped, overflow, silent clamping). If the FastPath CMSHeap can only hold integer counts, consider keeping the f64 path for SUM or rejecting non-representable inputs loudly.

[milindsrivastava1997]

Fractional SUM values lost. Every sample is rounded to i32 and anything ≤0 is dropped by update(). topk(5, sum by (pod)(cpu_seconds)) with gauge samples like 0.3/0.45 → all round to 0 → empty heap, empty result. 1.4 and 1.6 count as 1 and 2, skewing the ranking. The old f64 path got this right.

[milindsrivastava1997]

Silent clamping. value.round() as i32 saturates anything above i32::MAX (e.g. a 5e9 bytes sample → 2147483647) and maps NaN (staleness marker) to 0, which is then dropped. Both are wrong results with no signal — violates "Fail loud" (code-design checklist §1).

[milindsrivastava1997]

Stale doc comment (~line 729, outside the diff): the updater doc still says it wraps CountMinSketchWithHeap and feeds "the sample value itself", but it now rounds to i32 on the FastPath CMSHeap. Checklist §3: docs must change with behavior.

[milindsrivastava1997]

Memory undercounted. memory_usage_bytes counts only the counter matrix; it ignores the heap (up to heap_size owned String keys plus slot/position indexes). With long semicolon-joined label keys this can rival a 3x1024 i32 matrix, and halving the counter estimate widens the gap for the memory budget.

[milindsrivastava1997]

i32 counter overflow. Vector2D adds with plain +=. Once a key (plus its hash collisions) exceeds 2^31 (~2.1 GB of netflow bytes), debug panics and release wraps negative → negative min-estimate, heap evicts the real heavy hitter, query_key returns a negative value. CountMin::merge (increment_by_row) overflows the same way. The old f64 counters had no such limit.

[milindsrivastava1997]

Non-positive counts silently ignored. update() drops counts ≤0 with no error, so e.g. negative deltas from counter resets disappear and the sketch undercounts while looking healthy. Checklist §1: "Don't let a failure path quietly resolve to a default, an empty value, or a swallowed error."

[milindsrivastava1997]

Panics on non-string HeapItem, and the match is duplicated in get_topk_keys and serialize_to_json. If another construction/merge path ever puts a non-string key in the heap, a query or serialization crashes the engine thread. Suggest one helper returning Result.

[milindsrivastava1997]

serialize_to_bytes now always returns Err, and deserialization was removed. Any sink/store calling SerializableToSink::serialize_to_bytes on a boxed top-k accumulator (Kafka/msgpack output, nested serializers) now fails on every top-k window; previously it produced msgpack.

[milindsrivastava1997]

Merge result depends on order. Merged heap capacity is now the first accumulator's size instead of min of both: merging (5, 3) gives 5, (3, 5) gives 3, so different window-merge orders can yield different top-k sets. The rewritten test asserts 5 where the old one asserted 3, which hides the change.

Also: the rows/cols pre-check exists only because CountMin::merge panics on mismatch. That validation belongs in sketchlib as a Result-returning merge so other callers are protected too; the new error message also drops the actual dimensions the old one reported.