feat: sketch capability model + per-stage budgets + optimization problem formulation

[zzylol]

Summary

Wires the existing DeploymentConstraints into the actual optimizer pipeline.
Previously DeploymentConstraints and with_constraints() were defined but
never called — the optimizer ran unconstrained.

Changes

controller/src/main.rs

• Both handle_plan() optimizer calls now use QueryOptimizer::with_constraints()
  instead of QueryOptimizer::new()
• DeploymentConstraints::from_budgets(&StageResourceBudgets) converts
  workload-derived budgets into optimizer constraints

controller/src/algebra/optimizer.rs

New: DeploymentConstraints::from_budgets()

• Converts StageResourceBudgets (agent/backend memory) into DeploymentConstraints

Enhanced: DefaultCostModel::estimate()

• WindowedAgg costed with actual sketch memory from directory::estimated_sketch_memory_bytes
• Memory penalty: sketches exceeding agent_memory_bytes get 10× memory cost
• Bandwidth penalty: output exceeding agent_backend_bandwidth get 10× bandwidth cost
• Sliding window penalty: if agent_supports_sliding is false, sliding WindowedAgg gets 10× CPU cost
• Partition and Dedup have non-unity cost factors (0.8, 0.9)

How it works

main.rs:
  budgets = StageResourceBudgets::from_workload_chars(&wc)
  constraints = DeploymentConstraints::from_budgets(&budgets)
  optimizer = QueryOptimizer::with_constraints(raw_bps, constraints)
  (opt_qe, _) = optimizer.optimize(qe)

The optimizer's cost model now returns inflated costs when constraints are
violated, causing cost-sensitive rewrite rules to prefer alternatives
(e.g., deferring a large sketch from Agent to Backend).

Tests

3 new tests:

• constraints_from_budgets — verifies conversion from StageResourceBudgets
• constrained_optimizer_penalises_large_sketch — tiny budget → 10× memory penalty
• unconstrained_optimizer_normal_cost — no constraints → normal cost

318 total tests pass.

🤖 Generated with Claude Code