Do not silently select raw fallback when logical candidate costs are unavailable

[milindsrivastava1997]

Problem

avg by (job) (data) exposes two separate failures in logical candidate selection. The semantic rewrite (sum by (job) (data) / count by (job) (data)) is generated, but the root silently resolves to raw KeepPreAsap / exact fallback.

Problem 1: absent costs become strategy-order selection

When alternatives have no usable comparable cost, selection falls back to candidate discovery order. The deployment registers SketchAlgorithmStrategy before SemanticEquivalentRewriteStrategy, so the former's Avg pass-through candidate wins over the latter's realizable sum / count rewrite. This is not an evidence-based choice.

Current effective behavior:

no comparable cost
  → preserve strategy registration/discovery order
  → pass-through appears before semantic rewrite
  → exact fallback is selected silently

Candidate resolution options

• Fail loudly (preferred for evidence-required planning): if a root has materially different legal alternatives and none can be compared, return a typed cost_unavailable planning error with the candidate identities and missing inputs.
• Explicit deterministic policy: document and surface a deliberate fallback rule, such as preferring a realizable summary/rewrite over KeepPreAsap. This is only valid if product policy accepts that choice without measured cost evidence.
• Keep discovery order: not recommended; it makes registration order an undeclared optimizer policy.

Problem 2: ControlPlaneCostModel supplies no root-alternative cost

select_workload_impl calls Planner selection using ControlPlaneCostModel. That model currently does not implement a finite estimate_cost / candidate_cost for these logical root alternatives, and inherits the CostModel trait's NaN placeholder. The selection explain trace serializes this as:

"estimated_cost": null,
"estimated_cost_status": "not_reported_by_cost_model"

The backend has some later-stage lifecycle/window evidence, but it does not currently produce a complete comparison between:

raw exact fallback
vs.
maintain/read Sum + maintain/read Count + final divide

Candidate resolution options

• Fail loudly (preferred until complete costs exist): have ControlPlaneCostModel::candidate_cost return unavailable for unpriced root alternatives, and make selection fail with a typed diagnostic rather than treating NaN as a tie.
• Implement complete-plan costing: provide finite, comparable costs for the raw baseline and every reachable rewritten candidate, including maintenance, readout, final exact operations, query/evaluation rate, and selected window framework. This is the long-term fix.

Acceptance criteria

• Candidate selection never silently treats non-finite/unavailable cost as an equal-cost tie.
• The trace identifies the missing cost inputs and the policy/error that resolved the group.
• A regression case for avg by (job) (data) proves that unavailable costs produce the chosen typed error or an explicitly documented policy outcome.
• If complete costs are supplied, the selected root is demonstrably the lower-cost complete plan.

Related: #702 (root cause #3).

[milindsrivastava1997]

Why complete workload quotes cannot currently choose the rewrite

The avg by (job) (data) rewrite is not later overwritten by physical planning; it is discarded earlier, during logical root selection.

Current sequence:

1. Planner considers two logical roots
   - raw `avg by (job) (data)` pass-through
   - rewritten `sum by (job) (data) / count by (job) (data)`

2. Logical selection has no finite comparable root-alternative cost.
   It falls back to strategy/discovery order and commits raw pass-through.

3. Physical candidate enumeration starts only from that committed raw root.

4. `workload_cost_evidence` quotes those physical candidates and selects among them.

Therefore existing workload_cost_evidence cannot make the rewrite win: it is consulted only after the rewrite no longer exists in the physical candidate inventory.

The desired sequence is:

1. Keep both legal logical roots.
2. Realize/enumerate physical candidates for both roots.
3. Quote every complete physical candidate via `workload_cost_evidence`.
4. Select the lowest-cost complete plan across raw and rewritten roots.

This also clarifies the two failure modes in this issue:

1. When logical root costs are unavailable, strategy order silently decides the result.
2. ControlPlaneCostModel does not supply a finite root-alternative cost for this rewrite, while complete snapshot quotes arrive too late to participate in root selection.

An interim evidence-required policy could fail with a typed unavailable-cost error at step 2. The long-term solution is to move the decision to step 4, where complete plan costs can be compared.

[milindsrivastava1997]

can you please check this @zzylol ^