Preserve Planner aggregate semantics through maintenance and serving

[zzylol]

Why

Fixes #701 and #702. Reported PromQL workloads need correct window populations, executable aggregates, and whole-expression accuracy contracts.

What

Support shared current-series quantile/TopK/sum/count/average, typed minimum, temporal average, and guarded quantile ratios. Correct both pane and full-window reads while retaining native fallback.

How

Pin Planner PR ProjectASAP/ASAPPlanner#404 at 8a943abc. Planner rules construct typed candidate DAGs; the compiler lowers their operators and compares complete physical cost quotes. Runtime enforces current-value membership, window coverage, and division domains.

Before this PR

Instant aggregates could fail compilation, use temporal populations, or remain native-only. Minimum and temporal average lacked usable lowering. Sliding reads could merge overlapping snapshots or a preceding pane. Quantile division lacked a sufficient expression certificate.

After this PR

Mixed dashboards can maintain temporal summaries alongside shared current-series populations, including replacements, staleness, expiry, and maximum-k reuse. Ten quantiles share compatible state. min_over_time and avg_over_time execute from typed state; checked DDSketch division sizes operands for the requested expression bound and falls back outside its domain.

Evidence

A production-HTTP workload checks 46 queries at two successive evaluations against fresh Prometheus 3.5.0, including 15m/5m windows, multiple cadences, grouped instant aggregates, and both ratio forms. Values, labels, 1% quantile/ratio tolerance, and two zero-denominator fallbacks pass. The current-series process additionally compares 20 query forms through replacements/staleness.

See the developer review. Performance measurements and screenshots: not applicable; correctness quotes are synthetic.

Verification

• Unit tests: 2,098 backend workspace library tests passed. Focused pane regression also passed after the final range filter.
• End-to-end tests: 14 compatibility process tests passed; one existing Collector-schema test remains ignored. Includes real Prometheus comparison and durable-summary restart.
• Other checks: workspace/all-targets clippy with warnings denied, affected data-plane checks after final edits, formatting, and diff checks passed.

Regressions cover typed minimum/average/ratios, equal-valued series counts, shared current membership, invalid division domains, pre-epoch window admission, full-window pending/empty identity, and exclusion of preceding panes.

Architectural decisions

Semantic rewrites stay in Planner. Current-series state retains exact live values; it is not an append-only DDSketch. A 1% ratio requires component accuracy below approximately 0.4975%, not merely a shared 1% sketch. Deployment still requires complete cost evidence.

Limitations and follow-up

Depends on Planner PR 404. The boundary-aligned native oracle requires Prometheus 3.x semantics; 2.x includes the left boundary. Unsupported domains/coverage retain native fallback. Collector schema compatibility and measured performance benefits remain separate work.

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: