feat(planner): retain evidence-dependent candidates without certifying them

[zzylol]

Why

Planner #454 needs to show constructible alternatives even when external facts are absent, while keeping unknown accuracy, cost, and runtime support distinct from a proven deployment choice.

Before this PR: a grouped Count with no Hydra shared-grid evidence, or a Count-ranked TopK without interval evidence, lost candidates during construction or root-target filtering. A mixed exact/summary operation could also be selected from a permissive default capability flag. A legacy NaN estimate could appear as an available candidate cost.

After this PR: those logical candidates remain in PlanSpace with symbolic unknown guarantees; direct DDSketch ratios remain visible under the existing #449 behavior. Known-invalid evidence is rejected (including a typed Hydra rejection). Unknown mixed-execution support remains visible but is not automatically selected. Non-finite/negative default costs are unavailable, with a separate explicit legacy qualitative-ranking policy. Backend still owns physical binding and final deployment acceptance.

Changes

• Carry unknown Hydra, TopK, HLL, and relative-composition guarantees through target-aware search; reject already-impossible targets using only an optimistic floor, never as certification.
• Expose candidate-level accuracy and runtime-support status; gate automatic composition selection and preserve exact materialization fallback.
• Audit cost availability, document all Planner-side evidence gates and the backend boundary, and cover no/partial/invalid evidence plus export and end-to-end selection.
• Track backend integration in Integrate Planner evidence-dependent candidates into backend selection ASAPQuery-backend#752.

Verification

• cargo test --workspace --quiet
• Focused asap-aware-mapping and exact_composition tests after the final edits
• cargo fmt --all and git diff --check

Addresses #454.

[milindsrivastava1997]

Is it possible to have some before vs after e2e examples to show the difference?

[zzylol]

Added a before/after table and reproducible PromQL-to-Planner checks in docs/design_docs/architecture/evidence-dependent-candidates.md (commit 025c88b). The two new differences are: count by(job)(up) now retains two uncertified Hydra alternatives with no shared-grid evidence, and entropy_over_time(m[5m]) now retains its uncalibrated readout (guarantee: None). Neither becomes an automatic selected plan. I also included the quantile ratio as a control: its evidence-free candidate already existed after #449, so this PR does not claim that change. The tests start from PromQL text and verify the candidate/selection states; they do not execute a deployed backend query.