Module ranker 

Optimizer ranker (Picker 1) substrate — Phase 1.1 of the picker-work arc.

Builds the types + filter-chain pipeline the in-flight optimizer ranker will operate against. No SystemTopology integration, no candidate enumeration, no cost composition, no filter implementations — all of that lands in subsequent sub-phases:

Sub-phase	Scope
1.1	This module — AlternativeSet, Candidate, AlternativeFilter, chain
1.2	Candidate enumeration via SystemTopology (cross co-located backends)
1.3	Precision + tolerance hard filters (PrecisionGuarantee consumers)
1.4	Cost ranking — Layer-1 static composition
1.5	ExecutionPlan carries AlternativeSets; compile_plan integration

See docs/session-prompts/phase-1-1-alternative-set-filter-chain.md for the TDP resolutions and the full Phase 1 plan.

The two pickers

Per the 2026-05-30 picker-alternatives audit, the architectural endpoint is two pickers with very different scopes:

• Picker 1 — the optimizer ranker (this module’s eventual consumer). Plan-time machinery that enumerates candidates via SystemTopology, filters by hard correctness constraints, ranks survivors by composite cost (Layer-1 static + Layer-2 Judge data), and preserves the top-N per architecture v1.0 §04.
• Picker 2 — the runtime selector. Dispatch-time component that reads pre-blessed AlternativeSets and selects among the top-N based on layer-3 telemetry, checking SystemTopology generation before each dispatch chunk.

Picker 1 lands across Phases 1–3; Picker 2 is Phase 5.

Re-exports

pub use alternative_set::AlternativeSet;

pub use alternative_set::DecisionContext;

pub use alternative_set::KEEP_PER_DEVICE;

pub use candidate::Candidate;

pub use candidate::CouplingAdjustment;

pub use chain::apply_filter_chain;

pub use chained_selector::ChainedSelector;

pub use cost::CapabilitiesLookup;

pub use cost::TransferEstimator;

pub use cost::apply_inbound_transfer_costs;

pub use cost::composite_ns;

pub use cost::compute_static_costs;

pub use cost::default_backend_rates;

pub use cost_vector::AccuracyClass;

pub use cost_vector::CostVector;

pub use cost_vector::MemoryTiers;

pub use cost_vector::PrecisionDigits;

pub use device_load::DeviceLoadSelector;

pub use device_load::LOAD_TIER_IDLE;

pub use device_load::LOAD_TIER_MODERATE;

pub use device_load::LOAD_TIER_SATURATED;

pub use device_load::LoadLookup;

pub use device_load::MockBackendStreams;

pub use device_load::MockCombinedRuntime;

pub use device_load::load_tier;

pub use device_load::load_tier_for;

pub use device_load::mock_combined_lookup;

pub use device_load::mock_load_lookup;

pub use enumerate::enumerate_candidates;

pub use filter::AlternativeFilter;

pub use filter::FilterClass;

pub use filter::FilterContext;

pub use filters::AccuracyBound;

pub use filters::AccuracyPreferenceFilter;

pub use filters::DEFAULT_ACCURACY_BOUND;

pub use filters::PrecisionFloorFilter;

pub use filters::PrecisionRequirement;

pub use filters::StridedInputPreferenceFilter;

pub use filters::default_chain;

pub use judge::HashMapJudge;

pub use judge::JudgeOracle;

pub use judge_aware_selector::JudgeAwareSelector;

pub use placement_dp::ChainInput;

pub use placement_dp::PlacementDp;

pub use route_picker::LOOKAHEAD_K;

pub use route_picker::RouteCache;

pub use route_picker::TelemetryFingerprint;

pub use route_picker::pick_route;

pub use route_picker::resolve_branch;

pub use runtime_selector::RuntimeSelector;

pub use runtime_selector::WinnerSelector;

pub use vram_pressure_selector::BackendRuntimeHandle;

pub use vram_pressure_selector::BackendRuntimeLookup;

pub use vram_pressure_selector::OutputBytesEstimator;

pub use vram_pressure_selector::VramPressureSelector;

pub use vram_pressure_selector::default_estimate_output_bytes;

Modules

alternative_set

AlternativeSet — bounded collection of Candidates at one graph decision point.

candidate

Candidate — one (kernel, placement, precision, cost) bundle that the optimizer ranker is considering at a graph decision point.

chain

Filter-chain pipeline: applies an ordered list of AlternativeFilters to an AlternativeSet, respecting each filter’s FilterClass.

chained_selector

ChainedSelector — the production Picker 2 composition: VRAM-pressure GUARD → Judge-measured RANK → static-winner fallback.

cost

Cost composition + ranking for the picker. Phase 1.4 of the picker-work arc.

cost_vector

CostVector — the per-path cost VECTOR the rankers produce, and the Pareto-dominance relation over it.

device_load

DeviceLoadSelector + the load_tier bucketing — Step E Phase C, PR C2: the live-load arm re-pick.

enumerate

Candidate enumeration — walk the binding table for each (BackendId, DeviceLocation) placement the caller hands in and produce an AlternativeSet populated with one Candidate per registered alternative.

filter

AlternativeFilter trait + FilterClass + FilterContext.

filters

Concrete AlternativeFilter implementations shipped in Phase 1.3 of the picker-work arc.

judge

JudgeOracle — abstraction over the Phase 6b empirical profiler for the cost composer’s Layer-2 refinement.

judge_aware_selector

JudgeAwareSelector — Phase 5.2 runtime selector that re-queries the JudgeOracle at dispatch time and re-ranks candidates by fresh empirical measurements.

placement_dp

Carry-forward placement DP — planner Stage 3 (docs/session-prompts/load-time-incremental-planner.md, architecture 04-optimization §Load-time incremental planning, design pillars 3+4).

profile_oracle

ProfileJudgeOracle — the concrete JudgeOracle over the Judge’s measured profile data.

route_picker

The runtime route picker (“Picker 2”) — selects an arm at each Op::Branch decision point by live telemetry.

runtime_selector

RuntimeSelector — Picker 2 substrate.

vram_pressure_selector

VramPressureSelector — Phase 5.2 runtime selector that demotes or skips candidates whose backend can’t fit the projected output at dispatch time.