[milindsrivastava1997]

Why

main has been red since #1144 because of one test: the DAG viewer's test_a_plan_without_a_rate_fact_names_the_missing_fact. Since the MILP became Stage 3's search, a workload whose source has no rate fact fails with a bare "infeasible", and the user can't tell which fact is missing. The old engine named it.

What / How

Candidates that asap-cost can't price are dropped before the MILP, and their pricing error is kept in CandidateData::usage. When the MILP is infeasible, selection_error now appends each distinct pricing error to the Infeasible rejection's reason.

Before this PR

plan selection: no valid candidate: [Rejection { id: "workload", cause: Infeasible,
  reason: "no plan meets every query's accuracy target, latency bound and the reader
  constraints together (the selection MILP is infeasible)" }]

After this PR

... (the selection MILP is infeasible); candidates left out because they could not be
  priced: pricing source m needs an arrival rate (the arrival's rows_per_second, or the
  source's series_count with sample_interval), and none is declared

Also: unbreak main

main does not compile: #1164 imports crate::AnalyticalCostError in selector.rs, and #1187, merged before it, moved that import into lib.rs's test module. The second commit imports it from asap_cost::analytical instead.

Tests

• New regression test select::tests::an_unpriceable_workload_names_the_missing_fact: a workload with no source facts names the missing fact. It fails before the fix and passes after.
• The DAG viewer's test_server.PlanEndpointTests passes locally.

Key code interfaces added / modified: None (the private selection_error now also takes &MilpInputs).

Major algorithm changes: None.

🤖 Generated with Claude Code

[milindsrivastava1997]

Superseded by #1188, which landed the same pricing-reason fix and the selector.rs import fix.