scientific(coverage): predeclare recovery-fit numerical design before the 10k run #754

Finding

The prospective coverage design and simulation scenario are now versioned/fingerprinted, but analysis_engine::coverage_calibration_execution still creates the truth-K estimator configuration from analysis-local constants: seeds [7, 11, 19], maximum iterations 2_000, tolerance 0.001, plus the generic FittedCandidateKConfig ADR defaults. Those choices materially affect convergence, the numerical-failure denominator, fitted covariance, and interval coverage.

#694/#706 ensure that scientific recovery fits match the FittedCandidateKConfig supplied for a run, but they do not prospectively identify which FittedCandidateKConfig the 10,000-DGP coverage experiment must use. A future executor can therefore change initialization seeds, convergence budget/tolerance, or recovery hyperparameters while retaining the same tepp.coverage.nominal95.v1 validation fingerprint and tepp.simulation.rolling_origin_coverage.v1 scenario fingerprint. Exact source_head would reveal that code changed, but the experimental numerical design would not be a separately versioned owner contract fixed before outcomes are observed.

First owner repair

Introduce a model_selection-owned coverage-calibration fit design that:

• has a stable versioned identity distinct from validation and simulation identities;
• declares the current candidate strategy as one truth-K candidate per rolling-origin fit;
• owns the exact estimator seeds [7, 11, 19], 2,000-iteration budget, tolerance 0.001, prior variance 1.0, relation strength 0.25, ridge 0.01, topic smoothing 0.05, and step size 0.2 explicitly rather than inheriting mutable generic defaults;
• constructs the exact FittedCandidateKConfig for a supplied truth K and rejects invalid K through the existing model-selection owner;
• exposes a deterministic domain-separated SHA-256 fingerprint for the concrete truth-K configuration using canonical integer/string lengths and exact binary64 bits;
• pins the v1 fingerprint in a public contract and proves the produced FittedCandidateKConfig retains the declared seeds/iteration/tolerance.

This is the foundation prerequisite. The next consumer repair must make coverage_calibration_execution consume this owner instead of its local constants and bind the fit-design identity/fingerprint into shard/final evidence before the 10,000-DGP run. Do not copy this scientific truth into validation or simulation owners.

Boundary

This does not execute the 10,000-DGP study, change estimator arithmetic, define numerical-failure acceptability, or promote #680. model_selection owns the numerical fit design; analysis_engine remains the composition owner; validation_core remains the coverage criterion/evidence arithmetic owner; tepp_simulation remains the DGP owner.

Refs #639 #680 #694 #706 #733 #753.