fix(planner): model count-min sketch heap costs

.design_docs/optimizer-v1-implementation-plan.md

@@ -270,24 +270,24 @@ Allow a config with labels ⊇ query labels to serve that AQE. This is what enab
 - ✅ Plug real `AtomicCosts` values into cost model — done for CMS/HLL/KLL, see 2f above and
   "Running with real sketch-bench costs" below.
 - ✅ `CountMinSketchWithHeap` wrapper in sketch-bench (`cms-heap` family — sketch-bench PR #104,
-  closed sketch-bench#95) — but not wired to ASAPQuery's cost lookup yet, see status below.
+  closed sketch-bench#95) — wired to ASAPQuery's temporary reference cost model below.
 - ~~Add cardinality to sweep grids in `sketch-bench`~~ — decided unnecessary: CPU/mem costs for
   CMS/HLL/KLL are functions of structural params (depth×width, lg_k, K), not cardinality: only
   `CountMinSketchWithHeap` is cardinality-dependent, and that's the analytic-bound case below, not
   a sketch-bench sweep axis.
 
-**Status as of 2026-08-28:**
+**Status as of 2026-09-03:**
 1. `export_atomic_costs.sh` doesn't sweep `cms-heap` yet.
 2. Mapping gap: sketch-bench's `CmsHeapParams` is `{rows, cols}` only — `top_k` is a compile-time
    constant, Frequency-vs-TopK is two registry variants, not a param. ASAPQuery's grid sweeps
    `{depth, width, heapsize, count_events}` — `heapsize` has no home in sketch-bench's data at
-   all, confirming it must stay analytic. Unresolved: which of the 4 sketch-bench variants
-   (FastPath/RegularPath × Frequency/TopK) matches what ASAPQuery's `CountMinSketchWithHeap`
-   accumulator actually deploys, and what `count_events` maps to.
-3. Once (2) is resolved: extend `sketch_bench_key()` in `atomic_costs.rs` to translate
-   `CountMinSketchWithHeap` → the right variant + `{rows,cols}`, for CPU costs only.
-4. Analytic memory bound (`heap_size · avg_key_size`) still not implemented — needed regardless
-   of (1)–(3), memory was always meant to be analytic here, never a lookup.
+   all, confirming it must stay analytic. Resolved for the current runtime: use
+   `cms-heap-topk-regularpath-vector2d`; `count_events` is intentionally ignored because it
+   does not change the structural cost model.
+3. ✅ `atomic_costs.rs` translates `CountMinSketchWithHeap` to the selected `{rows,cols}`
+   reference row and scales CPU costs linearly from the fixed top-k=32 benchmark.
+4. ✅ Analytic memory bound: runtime CMS counters use 8-byte values, and heap memory uses
+   explicit average-key-size and per-entry-overhead assumptions.
 5. sketch-bench#14 (HLL register-width memory bug) — still open, still un-fixed.
 6. asap_sketchlib#69 (`subtract` unimplemented) — still open, `subtract_cpu_secs` absent by design.
 
@@ -323,7 +323,8 @@ Takes the same `ControllerConfig` YAML format as `asap-planner --input_config`
 (with a `metrics:` hints block for label schema — no live Prometheus needed).
 Prints deployed streaming configs and query configs to stdout. `--rho` is the
 placeholder arrival rate (see TODOs below — not real yet). `--atomic-costs` is
-optional; omit it and every candidate costs at the flat stub, same as before #549.
+optional; omit it and ordinary unbenchmarked candidates use the flat stub, while
+CMS-with-heap candidates warn and are dropped until a matching reference row is available.
 
 ### Running with real sketch-bench costs
 
@@ -350,9 +351,9 @@ cargo run -p asap_planner --bin asap-optimizer-cli -- \
   --atomic-costs path/to/atomic_costs.json
 ```
 
-`candidate-gen-dump`'s output labels each params row `[real]` (resolved from the table) or
-`[stub]` (fell through to `AtomicCosts::default()` — either an unbenchmarked family, or a
-benchmarked family's param point missing from the table).
+`candidate-gen-dump`'s output labels each resolved params row `[real]` or `[stub]`; candidates
+whose required cost row is missing are shown as `DROPPED`. CMS-with-heap uses its fixed-top-k
+reference model when the `{rows,cols}` row is present and is dropped otherwise.
 
 Wire-in decision (deferred): once Phase 3 (MIP + feasibility + label superset
 matching) lands, swap `Controller::generate()` to call `run_mip_pipeline()`
@@ -372,7 +373,7 @@ instead of `generator::generate_plan()`, likely behind an opt-in flag first.
 | `translator.rs` | `retention_count_for_assignment(Subtract)` | Returns hardcoded `1`; should be the actual checkpoint count needed to cover the full lookback |
 | `promql/generator.rs` | `generate_plan()` doc comment | Flags that `Controller::generate()` still uses the hardcoded path, not the optimizer — see "Offline Testing" section above |
 | — | Accuracy constraint | No `Error(a,g) ≤ ε_a` check exists anywhere; nothing stops picking an under-provisioned sketch (Phase 3d) |
-| — | sketch-bench | No cardinality sweep, no `CountMinSketchWithHeap` wrapper; `AtomicCosts` are still stub numbers (Phase 3c) |
+| — | sketch-bench | `export_atomic_costs.sh` does not yet sweep `cms-heap`; CMS-with-heap currently uses the temporary fixed-top-k reference model and must fail louder once swept coverage exists |
 
 ---