Preserve supported objective axes during Pareto pruning #723

Summary

Evaluation removes candidates using fewer axes than the objective function supports. A profile preferred by a quota, latency, rework or review objective can disappear before scoring.

Evidence and reproduction

On main fa0c79b9ae38eca2c59b2b1911becfe1a6b3c9ab, src/eval_harness/issue26.rs:421-464 normalizes and scores cost, tokens, wall time, churn and review, but it only scores profiles retained by the frontier. Its dominance test at lines 752-769 uses monetary cost and raw quality alone.

With equal raw quality, let A cost 1 and consume 100 tokens/time/churn/review units; let B cost 2 and consume 1 of each. The harness prunes B because A costs less, so a 100%-latency or 100%-quota objective cannot choose B.

The newer src/evaluation.rs:3126-3159 dominance path includes cost/quota/latency/quality, but omits retry/rework and human-review metrics that scoring supports at lines 3194-3266.

Acceptance criteria

• Dominance must include every typed raw axis used by the applicable objective, or defer pruning when that objective uses an unrepresented axis.
• Preserve raw quality safeguards and treat missing metrics as unknown rather than zero.
• Add independent trade-off fixtures for quota, latency, retry/rework and human review, showing that a fully weighted objective can retain and select its superior profile.
• Preserve deterministic ordering and rescore provenance across both evaluation paths.

Scope and verification

Source-level counterexample; no benchmark or provider run was performed. Closed #26/#150 delivered the harness/objective foundations. No current Issue/PR covers this pruning mismatch.