Final TODOs for Greedy optimizer

[milindsrivastava1997]

Non-empirical gaps in the Phase 2 greedy solver (asap-planner-rs/src/optimizer/), separate from plugging in real numbers:

• Costs are param-invariant: ingest_cost/query_cost ignore sketch params (CMS depth/width, KLL K, HLL precision, heap size) — the whole candidate_gen parameter grid is cost-invisible today. (Optimizer: param-aware atomic costs from sketch-bench #524)
• No accuracy/error-bound modeling: SketchProperties tracks mergeable/subtractable/subpopulation_aware but nothing about error guarantees, so greedy can pick the smallest param combo with no accuracy constraint. (Optimizer: CMS accuracy constraint (ACC) using accuracy_sla #526)
• subpopulation_count hardcoded to 1.0 in both cost functions — subpop-aware vs non-aware currently behave identically. (Optimizer: real N_g via Prometheus profiling, hints fallback #525)
• No cross-AQE sharing (facility-location) — greedy assigns every AQE independently, deferred to the Phase 3 MIP.
• Multi-statistic AQEs (e.g. avg = Sum+Count) unsupported — always falls back to EXACT.

For empirical calibration (tracked separately/later):

• Per-config atomic costs as functions of sketch params (not flat constants) — from sketch-bench. (Optimizer: param-aware atomic costs from sketch-bench #524)
• CostWeights from real cloud $/byte-sec and $/cpu-sec.
• arrival_rate_hz per (metric, spatial filter) from Prometheus scrape interval × series count.
• subpopulation_count per config from Prometheus series cardinality. (Optimizer: real N_g via Prometheus profiling, hints fallback #525)