Struct FrontierConvergenceOptimizer 

pub struct FrontierConvergenceOptimizer;

The PR-B2 per-device-frontier optimizer, re-expressed as a registered Optimizer. It runs after each pathfinder (lock-step) and is the in-graph MERGE/DISCARD counterpart of AlternativeSet::retain_per_device_frontier, which is applied per kernel-bearing node inside compile_plan.

Its two responsibilities:

1. Duplicate-path convergence (the trivial MERGE optimizer). Detect any two arms of a just-emitted Op::Branch that are forward-identical — same op, same inputs, same target backend. Two pathfinders (or one pathfinder firing twice) could in principle propose structurally-identical routes; the constitution requires “convergence merges only forward-identical paths”. The PR-A4 seed proposes one runner-up per fork, distinct from arm-0 by construction (different target_backend), so today no duplicate arm ever arises and this detection is a no-op — it is the trivial second optimizer the spec calls for, in place (with a debug assertion that it stays a no-op) so the contract is documented and enforced before a future multi-runner-up pathfinder can leak a duplicate arm into a branch.

2. Invariant guard (the never-strand / no-active-cycle contract). Assert that every Op::Branch keeps ≥1 viable (device, backend) route — never strand the last path (upheld by the per-node frontier inside compile_plan, re-asserted here over the in-graph form) — and that no branch the optimizer would touch is in ctx.cycle_guard (the incremental-reopt contract — batch optimize has an empty guard, but the guard documents and enforces the rule for the Phase-C runtime).

Trait Implementations

impl Optimizer for FrontierConvergenceOptimizer

fn name(&self) -> &'static str

Stable, human-readable name. Shows up in debug traces.

fn prune(&self, graph: &mut Graph, ctx: &OptimizationContext<'_>) -> Result<()>

Merge/discard candidate paths in graph. Runs after each pathfinder over the region that pathfinder just touched.