feat(planner): bind offline frequency evidence and preserve o11y fallbacks

[zzylol]

Before this PR

The control plane could not use query-matched offline frequency measurements to choose a configuration. Bare selectors, sort roots, and comparison/filter roots could also fail binding instead of preserving their exact query semantics.

After this PR

• Accept validated offline evidence for candidate ordering and fixed-snapshot frequency comparison. Require compatible parameters and formal accuracy minima; exact winners or unavailable evidence preserve exact execution.
• Preserve unsupported o11y root shapes as explicit exact fallback. Binding success remains separate from installed execution and query acceleration.
• Integrate current main, align all Planner dependencies and executable example snapshots with abb2f20e27091ac0c60715dc42b9b59c75b3447c, and adapt the typed physical-resource API while retaining the archived evidence JSON schema.
• Reject legacy keyed/non-column summary updates that the one-column runtime adapter cannot represent. In particular, constant unit weight must not be interpreted as sample-value SUM.
• Include offline replay/export examples and tests. Historical offline reports remain historical observations; they were not regenerated by this integration.

This PR supplies the offline evidence and binding layer. It does not establish end-to-end o11ybench execution, warm coverage, or performance/resource benefits versus Prometheus; those require the execution and matched-comparison follow-ups.

Validation

• Fresh-source compilation: 658 control-plane library tests, 6 offline-evidence tests, 1 exact-fallback integration test, and 14 data-plane legacy-summary tests passed.
• Both offline examples built against the pinned dependency. The pinned upstream corpus replay bound 27/27 queries in each of exact/default modes; this is binding evidence, not a serving or speedup result.
• Five Python export tests, package formatting, and diff checks passed.
• Shared-target source roots were explicitly refreshed before compiling to avoid reusing another worktree's test artifact.

[zzylol]

Documentation migration completed in c2a632e: docs/design_docs/empirical-o11y-execution-plan.md, docs/offline-o11y-final-2026-09-08.md and docs/user_guide/o11y-replay.md now live here, together with tools/empirical-bench. Cross-repository links were repaired; historical results remain explicitly historical. Planner PR ProjectASAP/ASAPPlanner#358 is closed in favor of this repository-owned documentation. Backend-only runtime work is separate and still under validation.

[zzylol]

Created the two stacked implementation follow-ups: #525 (cost-selected production execution replay) → #526 (matched exact comparison and scoped resource reporting). Both are Draft: conformance/regression tests pass, but the complete real-corpus run, matching provider calibration, isolated exact baseline and full lifecycle/resource accounting remain unverified. #525 also integrates the existing deployment foundation through #505; see its Stack section before merging. No performance-benefit or five-criteria-completion claim is made.