fix(data_plane): read delta-only sketch windows + count(HLL)

[zzylol]

Root cause (the NEXT layer behind #323/#324)

Both layers fixed earlier (#323 metric-name canonicalization, #324 reconcile + Ghost short-circuit) were necessary but not sufficient. The KLL quantile_over_time and HLL count(...) queries still returned an empty result tagged data_source: asap_query (never failing over to the archive). The router treats Ok(empty) as success and does not fall over, so the client saw "No result".

Two distinct, independent bugs caused the empty Ok:

1. Delta-stitching carry-in (the quantile bug)

The agent emits a periodic Full snapshot followed by many cheap Delta frames to save bandwidth. A short query window (e.g. [30s]) routinely contains only deltas — the Full landed earlier, outside the window. SketchStore::query_range's strict containment filter (epoch_columnar.rs: w.0 >= start && w.1 <= end) dropped that out-of-window Full, so the delta-apply reducer (cumulative_evaluate) could not establish a rolling base and returned Ok(None). The reducer then pushed a series row with an empty sample vector, and since any_window was true (windows were seen) it did not raise NoData → the engine returned Ok(empty).

Fix: query_range now splices in the most-recent Full snapshot ending at/before start as a carry-in base (new collect_ending_at_or_before on both epoch types). The reducer drops the out-of-range base from per-window output so it never leaks into the answer's time domain.

2. count(HLL) distinct-count idiom (the cardinality bug)

count(metric) lifts the outer count into both outer_agg = Count and required_capability = CardinalityApprox, while the bare-selector inner leaves the analyzer's trace function empty. The engine then (a) passed "" to SketchReducer::evaluate → UnsupportedFunction, and (b) re-applied the Count fold, which would collapse the cardinality estimate to the row-count (1.0).

Fix: the engine derives the reducer family from the typed required_capability when function is empty (effective_sketch_function), and suppresses the already-consumed Count fold for CardinalityApprox candidates (outer_fold_already_consumed). The HLL distinct-count is returned directly.

What this is / isn't

• It is a seal/unsealed read-path bug (carry-in) plus a query-shape-dispatch bug (empty-function + redundant fold) — NOT a reconcile/eviction bug. In-memory unsealed sketches are read correctly (matches the "可以in-memory, 不一定要seal" requirement).

Tests

• Re-enables the previously-failing controller_plan_to_query_full_roundtrip_hll e2e.
• New regression tests: index-level carry-in (range_query_carries_in_latest_full_before_window, range_query_no_carry_in_when_window_leads_with_full), engine-level quantile carry-in (quantile_over_time_kll_full_before_window_carries_in_base), the genuine delta-only-no-base data gap (quantile_over_time_kll_delta_only_no_base_is_empty), engine-level HLL count (execute_count_hll_returns_cardinality_not_rowcount), and analyzer shape (analyze_count_bare_metric_trace_shape).
• Full data_plane (726 lib + all e2e) and control_plane (771) suites green.

Possible further layers

• A delta-only window with no Full anywhere still returns empty Ok (no base to stitch). Today it is a silent empty rather than a NoData → archive failover; a follow-up could convert it to a capability-miss so the cold tier answers.
• The instant-vector projection takes the latest window's value, so an empty "watermark" frame as the freshest window yields 0; benign for liveness but worth confirming against the live watermark cadence.

Test plan

• Rebuild + redeploy the whole stack at this branch and re-run the 5 marquee queries (quantile_over_time KLL, request_size_bytes KLL, count(unique_users_per_min) HLL, plus the already-working sum-by and count cold-fallback).

🤖 Generated with Claude Code