Install coherent deployment plans with dataset-bound SDS identities

[zzylol]

Problem

Deployments need one coherent installed plan that binds queries and producers to stored state with known meaning. Routing fingerprints alone cannot distinguish equal metric expressions over different logical datasets, or represent two deployed outputs with the same semantics safely.

Before this PR

Backend configuration mixes plan concerns and does not establish the SDS identity contract from #737. A query binding cannot independently express the semantic computation and the authorized deployed output.

After this PR

Install a coherent PrecomputePlan, QueryPlan, transmission configuration and immutable catalog snapshot. Planner exports the persisted output's canonical typed dependency closure, including logical dataset identity; deployment binding connects it to a stored output.

Planner persisted-output semantics + logical dataset
    → SummaryDefinition / definition_id
    → deployed stored_output_id
    → installed writer and query bindings
    → validated stored records

Two KLL(latency) outputs named hot/rebuild may share a definition but are independently bound. Another dataset changes the definition; relocating the same dataset preserves it. Same-version restart restores valid state. A new plan version remains cold until fresh input is produced and never adopts previous-version payloads.

Catalog schema 6 and planning snapshot schema 3 establish this identity contract here. Typed Count/Rate support formerly in #771 is included because those semantics must remain distinct. Shared physical execution and its precompute execution design document continue in #774.

Obsolete installed-plan aliases, untyped projection decoders, the unused materialization wire adapter, and cross-version adoption metadata are removed. Recovery accepts only the current sidecar schema and propagates malformed or unsupported metadata errors. Producer partition rosters are explicit on the wire.

Dependency changes

• asap_sketch_codec contains the envelope encoding/decoding helpers used by backend ingest, DDSketch/KLL accumulators and wire-format tests. It supports removing the asap-precompute-rs Collector runtime dependency; it adds no sketch algorithms. The workspace member and data-plane dependency are intentional. With the Collector dependency removed, its Sketchlib patch block is unused and is removed too.
• Planner advances from cd7e9e0 to bccc837 because this implementation uses LogicalDatasetIdentity and SummarySemanticFragment::from_stored_output_in_dataset. Both APIs are absent at the old revision. All four Planner dependencies use the same immutable revision.

Validation and scope

The SDS implementation passed locally: 117 type-library tests, 431 control-plane library tests, 1,165 data-plane library tests, 8 HTTP API tests, 13 serving integration tests, the production-process restart/warm-up test, strict all-target Clippy and formatting.

Coverage includes dataset changes, endpoint relocation, independent hot/rebuild bindings, tampered definitions, writer/read consistency, same-version restart without re-ingestion, and new-version warm-up. Earlier downstream validation also passed Level 1, exhaustive synthetic Level 2 selection, 332 storage tests and 13 serving integration tests.

Discovery, calibration and workload replay now preserve version-3 dataset identity. Process fixtures use typed deployment configuration instead of removed flat aggregation documents. The 81 Python tests and seven focused sketch/component/monitor/forwarding process tests pass locally; the complete workspace and required CI are rerunning.

PR-specific evidence directories and generated archives are removed from the repository. Test summaries belong here; raw logs remain local or in CI.

Ad-hoc discovery, cross-version adoption and production-cost validation remain out of scope. Restricted native configuration helpers support explicit imported-state fixtures; production Planner compilation requires dataset identity. No manual deployment or human approval is claimed.