feat(sweeper): add heterogeneous P/D search [AIC-1772]

[jasonqinzhou]

Summary

• add independent prefill/decode model, system, backend/version, estimator/data, engine-control, scheduler, quantization, and KV-feasibility resolution with documented inheritance
• search backend pairs and role-specific parallel shapes under one GPU budget, retaining exact role estimator, engine-request, and candidate provenance
• support mixed vLLM/SGLang disaggregated replay, role-correct lowering, GPU accounting, legacy-calibrated rate matching, and categorized role failures
• fail closed for unsupported heterogeneous backend pairs, mismatched role identities/model architectures, invalid finite correction values, and adapters without explicit heterogeneous-P/D capability opt-in
• forward role systems paths and estimator database/transfer/FPM controls losslessly into the runner and Rust/Python engine compile path
• preserve the homogeneous input and replay path unchanged when no role override is configured

Stack dependencies

• Depends on refactor(perfmodel): unify estimator construction and CLI policies [AIC-1770] #13 (AIC-1770 estimator/data controls); exact stacked commit: 8fc904d
• Depends on feat(sweeper): preserve engine and request controls on canonical API [AIC-1771] #19 (AIC-1771 engine/request controls); exact upstream foundation c2671d0 is cherry-picked here as b4f7fbd
• This PR should be reviewed and merged after refactor(perfmodel): unify estimator construction and CLI policies [AIC-1770] #13 and feat(sweeper): preserve engine and request controls on canonical API [AIC-1771] #19. The base remains main so the complete dependency stack is visible.

Parity remediation

• require explicit prefill/decode backend-pair capability and retain deterministic considered/accepted/pruned diagnostics even when another pair remains viable
• keep RoleIdentity identical to the resolved per-role EstimatorSpec, including inherited backend-version fallback; reject incompatible model architectures before KV handoff
• forward configured role/shared systems_paths, performance database mode, transfer policy, FPM selection, and exact data roots through engine lowering
• prevent shared chunked-prefill controls from leaking into the decode request/provenance
• make all five disaggregated legacy correction knobs executable through a canonical replay correction spec and finite timing-scale validation
• reject standalone or effective per-role rate arithmetic overflow and validate every canonical role_rates value before serialization
• retain ordered counts plus pair/role/category pruning details on terminal NoViableParallelConfig.enumeration_reports when every heterogeneous pair is pruned, including explicit pinned domains

Exact-head validation (555f5695)

• pinned/enumerated all-pairs-pruned integration and search-space suites: 27 passed
• full root Python/Sweeper suite: 381 passed, 3 skipped
• exact CI Ruff scope: passed
• git diff --check: passed
• generator bridge diff audit: no files touched

The preceding remediation head also passed 456 workspace Rust tests. Earlier proportional legacy Python validation completed 3,578 passed, 10 skipped, 0 failures before a bounded stop; 66 selected slow build-tail cases were not completed locally.

Residual boundaries

• TensorRT-LLM disaggregated execution remains unsupported and is rejected by role/pair capability.
• decode-side chunked prefill remains disabled because it is a prefill-role scheduler control.
• optional adapters must explicitly advertise each supported heterogeneous backend pair after auditing role-specific candidate materialization.
• the 66 slow legacy build-tail cases remain for hosted CI; no failure was observed in the completed 3,588 outcomes.

Linear: AIC-1772

[linear-code]

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

Review in Linear