feat(delta): per-window delta producer for CMS/CountSketch/HLL (delta-against-empty)

[zzylol]

What

Extends the per-window delta-against-empty producer (landed for DDSketch in #456) to CountMinSketch, CountSketch, and HLL on both edge runtimes (asap-precompute-go and asap-precompute-rs). KLL stays full-only; DDSketch is unchanged.

Each window now emits a delta computed against an empty base, so the wire payload carries that window's own per-window state with no cross-window subtraction — the prerequisite for correct window-scoped backend aggregation.

How

asap-precompute-go (sketches/{cms,countsketch,hll}.go):

• Each wrapper implements DeltaAgainstEmptyBase(), returning the encoded empty-sketch snapshot the SnapshotCache caches as the outbound base after each window-close emit. The next window's ComputeDeltaAgainst (already wired to the sketchlib-go ComputeDelta for each family) then diffs against the empty base.
• CMS msgpack mode (which cannot carry deltas) returns a nil base to keep legacy always-refresh. SnapshotCache.ComputeDelta now treats a nil/empty empty-base as "fall through to legacy refresh", so a family can opt in conditionally.

asap-precompute-rs (src/sketches/{cms,countsketch,hll}.rs):

• Each wrapper implements delta_against_empty_base() and a real compute_delta_against that decodes the prior envelope and diffs via the asap_sketchlib compute_delta for that family (CMS/CountSketch cell deltas, HLL register delta), with a full-snapshot fallback on an empty/undecodable prior.
• apply_delta now dispatches on payload shape (full-state envelope vs raw delta proto via apply_delta_bytes), mirroring the DDSketch wrapper. Stale "always-full" caveats in sketches/mod.rs / precompute.rs doc comments updated.

The SnapshotCache dispatch was already sketch-agnostic (any sketch implementing the empty-base interface opts in; others fall back to legacy always-refresh), so no dispatch logic changed — only the per-family opt-in and the conditional nil-base fall-through.

Tests

Mirroring DDSketch's tests, for each of CMS / CountSketch / HLL:

• applying the emitted delta to an empty base reconstructs the window's state (matrix / registers / estimate);
• two consecutive windows each emit their own state (no cross-window subtraction). HLL's disjoint-window test pins that the empty-base reset defeats register-MAX leakage (delta-baseline-contract.md §1.5 / §2.3).

Go: sketches/{cms,countsketch,hll}_test.go. Rust: src/snapshot_cache.rs. Go go build/go vet/go test ./sketches/... and Rust cargo build/cargo test are green (Rust built against the asap_sketchlib branch carrying #59).

Byte-parity note: Go↔Rust per-window delta-frame byte parity is not directly asserted in this PR (no in-tree golden delta fixtures), but the underlying delta codec is the byte-parity twin verified by asap_sketchlib#59's Go-golden envelope probes; each Rust compute_delta is documented as byte-identical to the sketchlib-go proto.Marshal output for the same inputs.

Dependencies

• Depends on asap_sketchlib#59 (CMS/CountSketch/HLL compute_delta / apply_delta_bytes). The asap-precompute-rs path-dep keeps the normal relative path; build against an asap_sketchlib carrying feat: unify sketch processor config interface #58 (DDSketch) + fix(series): send full attrs until collector confirms series ID (multi-hop fix) #59.
• The backend base rotation (mvp_report: fix DDSketch family metric name (http_latency_ms → http_requests_total_latency_ms) #346) is already sketch-agnostic, so no backend change is needed.

🤖 Generated with Claude Code