feat(sweeper): integrate heterogeneous execution search [AIC-1772]

[jasonqinzhou]

Summary

• compose heterogeneous prefill/decode search from feat(sweeper): add heterogeneous P/D search [AIC-1772] #22 with the finite execution domains from feat(sweeper): safely expand explicit parallel domains [AIC-1773] #18
• enumerate role-specific GPU/worker, TP, PP, attention-DP, MoE TP/EP, and CP choices using each role's model, system, backend/version, estimator data, and engine controls
• enforce shared GPU budget, allowed replica sizes, combined replica ceiling, and per-role worker ceilings before either rapid or thorough sampling
• preserve actual batch/context choices, role sample identity, AIC PP/CP arguments, runner topology matching, pair capability gates, correction factors, and deterministic pruning diagnostics
• add an end-to-end candidate-count/materialization fixture proving asymmetric PP/CP/worker domains lower losslessly into native replay
• condition topology/backend feasibility on each concrete scheduler point so rapid projection and thorough enumeration share the same complete finite legal domain instead of applying the old 8192-token/256-batch defaults
• deterministically project rapid scheduler samples onto the nearest legal scheduler/topology pair, including pinned-topology branches, and retain requested/actual scheduler and topology metadata
• retain ordered aggregate scheduler-point pruning counts and diagnostics for partial branches and homogeneous all-pruned terminal reports
• propagate selected max_num_seqs into AIC KV-block estimation and validate MoE TP/EP topology against both flat generated arguments and nested public AIC timing configuration
• dispatch the correct infeasibility keyword for both seeded pyvizier trials and unseeded Vizier service-client trials

Dependencies

This is an integration-only stack and should not merge before:

• refactor(perfmodel): unify estimator construction and CLI policies [AIC-1770] #13 — estimator and performance-data controls
• feat(sweeper): safely expand explicit parallel domains [AIC-1773] #18 — parallel and execution search dimensions (integration base: e592d0b22b1b7776177c63935656e004d6cfa14b)
• feat(sweeper): preserve engine and request controls on canonical API [AIC-1771] #19 — engine and request controls
• feat(sweeper): add heterogeneous P/D search [AIC-1772] #22 — heterogeneous P/D search (integrated corrected head: 555f5695d53b673309fe7797d40a9e7a1b6a401f)

Validation

• uv run --project python/aisimulate --extra dev pytest -q tests/sweeper/test_parallel_projection.py tests/sweeper/test_search_space.py tests/sweeper/test_sampler.py tests/sweeper/test_heterogeneous_integration.py tests/sweeper/test_model_hw.py tests/test_aic.py tests/test_runner.py -p no:timeout — 130 passed
• uv run --project python/aisimulate --extra dev pytest -q tests -p no:timeout — 425 passed, 3 documented packaging skips
• Ruff check across all six follow-up Python files
• git diff --check
• Rust tree unchanged by the follow-up; prior integration validation at c517009 was 456 passed with cargo fmt --all -- --check green

Tracking

• AIC-1772

[linear-code]

AIC-1772 [SWEEPER] Add heterogeneous role-specific disaggregated search

Review in Linear