[milindsrivastava1997]

Part of #1067. The test-only reference pricer failed a whole selection on a candidate the search rejects.

Why

planner_layering_example2::production_memo_selects_a_validated_example2_plan failed intermittently in CI with plan selection selects: Cost { candidate: "E17-0", node: 4, error: Overflow } (#1166 now ignores it). The search rejects a candidate whose cost overflows (cost_not_representable); the reference pricer, which prices every memo choice of the selected round for tests and stage_pipeline's reference_pricing, returned the error instead. Which round is selected depends on timing, so the failure appeared only when that round held such a candidate.

What

The reference pricer rejects a candidate whose cost overflows, with the new RejectionCause::CostNotRepresentable, and selects among the rest.

How

select_among_plans skips an overflowing candidate when fitting the calibration and, when offering it, records a CostNotRepresentable rejection instead of returning ErrorSelectingPlan::Cost. The search's own overflow handling and PlanSelector::offer_before are unchanged.

Key code interfaces added / modified

crates/plan-selection/src/report.rs — RejectionCause stays, gains one variant:

pub enum RejectionCause {
    AccuracyTargetMissed,                 // unchanged
    CountMinWeightsNotProvenNonNegative,  // unchanged
    NoAccuracyGuarantee,                  // unchanged
    LatencyBoundExceeded,                 // unchanged
    LatencyOverRepeatInterval,            // unchanged
    OverClusterCapacity,                  // unchanged
    /// Its cost overflows the cost model's arithmetic, so it cannot be
    /// priced. Written by the reference pricer; a failed check.
    CostNotRepresentable,                 // new
    Costlier,                             // unchanged
    Slower,                               // unchanged
    MoreMemory,                           // unchanged
    TiedLostOnEnumerationOrder,           // unchanged
    Infeasible,                           // unchanged
}

Serialized as cost_not_representable; added to the DAG viewer's failed-check causes, its README, the stage_pipeline output docs and asap-aware-mapping-contracts.md.

Major algorithm changes

None.

Before this PR

When the selected round held a candidate whose cost overflows, the reference pricer returned Err(Cost { .. Overflow }), failing tests that read it (and stage_pipeline's reference pricing), although the search had rejected that candidate.

After this PR

That candidate is listed under the reference's rejections with cause cost_not_representable; selection among the rest proceeds.

Evidence

the_reference_pricer_rejects_an_unpriceable_candidate fails on main (the selection errors instead of returning NoValidCandidate) and passes with the fix. The CI failure itself is timing-dependent and not reproduced.

Verification

• Unit tests: plan-selection tests::the_reference_pricer_rejects_an_unpriceable_candidate — candidates whose cost overflows are each rejected as CostNotRepresentable, not a selection failure.
• End-to-end tests: devtools viewer_contract (2) and tools/dag-viewer stages.test.mjs (19) pass. Full workspace suite not run locally; relying on CI.
• Other checks: workspace clippy -D warnings (CI's flags), fmt.

Snapshot changes

None.

Architectural decisions

• Mirrors the search's rejection in the reference pricer instead of changing PlanSelector::offer_before, which the old engine uses, so no production path changes.
• Dropped from this PR: classifying a deadline passed during execution-scope expansion as a deadline stop. Since feat(plan-selection): group selection MILP over execution contexts as Stage 3's default search (#1095) #1144 the default search (group MILP) does not run that expansion; it would only affect ASAP_STAGE3_SEARCH=old.

Limitations and follow-up

• Tests that fail on timing but are not listed in Planner tests depend on wall-clock search budgets and flake under load #1067 are noted there.

Human review — do not complete with an agent

• The MVP boundary is correct.
• New conceptual layers or public interfaces are necessary.
• The before/after description matches the intended product behavior.
• Human reviewer:
• Decision and rationale:

🤖 Generated with Claude Code