fix(promql): bind temporal TopK update semantics on the shared workload plan

[zzylol]

Why

Temporal TopK must rank series by the selected producer's update semantics. A count contributes one per observation; a sum contributes the sample value. Losing this distinction, series identity, or materialization identity returns incorrect winners.

Before this PR

The branch used an obsolete Planner readout-weight API, conflicted with the workload migration, and a single-series fixture could not prove ranking correctness.

After this PR

For one window, api has four observations summing to 10, worker has two summing to 200, and cron has three summing to 30:

• topk(1, sum_over_time(m[5s])) returns worker=200.
• topk(1, count_over_time(m[5s])) returns api=4.

The next window changes the sum winner to api=26. Instant/range tests check winners, values, cardinality and timestamps. The existing backend item label contains canonical series identity; native Prometheus label reconstruction is not claimed here.

Implementation and dependencies

Based on #522; merge the #513–#522 stack first. Planner #356 and #357 are merged; this branch pins revision c3410d14865497758d212e4265ad25c782187de1.

• Derive heap weight_mode from SummaryAgg.input.weight; reject unsupported weights.
• Retain query-scoped membership evidence in canonical workload selection.
• Keep complete series identity for ungrouped heaps and distinct SIDs for distinct physical policies.
• Reuse fix(topk): preserve typed count updates and migrate acceptance to registered QueryPlans #511's materialization and multi-series process-test fixes; fix(topk): preserve typed count updates and migrate acceptance to registered QueryPlans #511 now contains its additional legacy-binding work.

Validation

• 656 control-plane library tests passed.
• 977 data-plane library tests passed with --test-threads=1. The first parallel run had five persistence failures; an isolated rerun and the serial suite passed.
• Two compatibility process tests passed, including exact shared dashboard and temporal TopK.
• Distributed production-process test passed with actual Collector Rust runtime.
• Select complete bound workloads with strict cost and capability evidence #522's exact-only quote-preparation regression and all 29 control-plane binary tests passed.

Self-review, not independent approval. Legacy sketch-envelope fixture failures tracked on #511 are not covered by the canonical-profile acceptance above. No merge performed.

Closes #502. Contributes the #706 rows of #501.

Planner #357 compatibility update

Synchronizes the latest #522 changes, including complete exact-fallback input-source costing, and pins merged Planner revision c3410d14865497758d212e4265ad25c782187de1. Retains the reviewed backfill fixes: policy-aware series IDs and shared live/replay TopK update semantics. Regenerate cost manifests and quotes for the new Planner revision.

Latest combined-stack validation with Planner #357: 658 control-plane library tests, 29 control-plane binary tests, one control-plane integration test, 978 data-plane library tests (serial), two compatibility process tests, and the distributed production-process test passed. An initial data-plane run exhausted the machine's root filesystem; the full rerun passed using a temporary directory on /mydata.