feat(sketch_algebra): realize Frequency (AggIntent::Extension) as a real Cms sketch

[zzylol]

Summary

Phase 2 of 5 toward supporting named-key SketchQuery::PointCount (count(cms_metric{item="x"})) — see the earlier phase (data_plane's SketchQuery::TopK support, #412) and the upstream prep (ASAPController#162, not yet merged).

Blocked on ASAPController#162 merging. This PR temporarily bumps control_plane's asap-ir/asap-l2/asap-sketch/asap-plan rev pins to e4cea66 (the feat/extension-realization-hook branch commit, ASAPController#162) so this can be built/tested now. Re-pin to the actual merged commit once #162 lands — same pattern as the fc09c3a → d4c1756 bump before ASAPController#161 merged.

AggIntent::Extension{ext_kind: "frequency"} (control_plane's point-frequency-via-CMS intent, count(*) WHERE key = k) has always mapped to Implementation::PassThrough — a documented, accepted regression (typed_binding_endpoint_request_freq_declines_pending_upstream_extension_support) since Step B of the plan-shaped-serving migration adopted asap_plan::bind::implement_tree_in_with without a local Frequency-binding workaround. ASAPController#150/#162 added the CostModel::{realize_extension, readout_extension} hooks needed to close this; this PR implements both on ControlPlaneCostModel:

• realize_extension("frequency", payload) parses the accuracy out of payload and returns Implementation::Sketch{kind: SummaryKind::Cms, params: Cms{width, depth}}.
• readout_extension builds SketchQuery::PointCount, falling back to {key: SampleValue, value: None} (the bare bucket total) until the actual filter value is threaded through payload — a separate, later change (Phase 3).
• ForcedFamilyCostModel (used by optimizer::rules::bind_workload_typed) delegates both new methods to its inner ControlPlaneCostModel, so the endpoint_request_freq contract row now commits instead of declining.

A genuine discrepancy found and resolved along the way: the legacy sketch_catalog::sketch_type_for_op's as_frequency arm maps Frequency to SketchType::CountSketch, while the newer, actively-tested capability_matching::pick_family/is_valid_pair maps it to SummaryKind::Cms — these are genuinely different sketch algorithms (SketchType has separate CountSketch/CountMinSketch variants), not a naming coincidence. Went with capability_matching's mapping since it's the tested, current source of truth for the typed binding path this change lives in; the legacy sketch_catalog-driven planner (a separate, still-coexisting code path, untouched by this PR) keeps its own CountSketch choice — not something this PR should or does reconcile.

Also updates emit::stage_config's SketchQuery::PointCount wire-JSON builder for the new value field (ASAPController#162).

Test plan

• cargo test -p control_plane --lib — 765 passed, 1 failed (the same single pre-existing, unrelated failure — invalid_sketch_type_override_falls_back_to_default — present on unmodified main too).
• 3 tests flipped from "declines pending feat(sketch_db): Phase 5 M2 — derive sketch sid from content hash #150" to "binds successfully": typed_binding_endpoint_request_freq_binds_cms (renamed), sketch_algebra::tests::frequency_extension_binds_cms (renamed), plus the CountMinSketch-override case in override_pins_nonquantile_family_over_misclassified_query and the two routing-table fixtures in emit::mod.rs.
• cargo test --workspace — same single pre-existing failure, nothing else.
• cargo build --workspace — clean.
• cargo clippy -p control_plane --lib --tests — verified every warning falls outside this diff's touched line ranges (all pre-existing).
• rustfmt scoped to touched files only.

🤖 Generated with Claude Code