Add query-conditioned observability for partial 4D gauges

[FlorianPfaff]

Purpose

Turn the recently added observable-subspace Sim(3) factor into a decision-relevant scientific kernel.

A rank-deficient overlap may locate points on its observed support while leaving an off-support physical query dependent on an unobserved gauge direction. Pooled point error and a favorable average covariance reduction do not identify that distinction. This PR projects a partial gauge likelihood through a declared downstream query and makes exact fallback possible before BayesianPhysTwin consumes the candidate.

This is the method-level bridge needed by #49 and the PointWorld/Flat'n'Fold path in #333. It does not open any protected target outcome.

Method

For a query Jacobian J_q in the centroid-normalized Sim(3) chart and a declared positive-definite output metric, the new kernel reports:

• the fraction of query sensitivity lying directly in the factor's observable subspace;
• the complementary sensitivity to the retained geometric nullspace;
• complete-prior query covariance before and after factor fusion;
• metric-weighted query-variance reduction; and
• the worst posterior/prior variance ratio over the supported query covariance.

Direct geometric support is deliberately separated from variance reduction transmitted only through prior correlation. A source-frozen three-criterion gate returns stable rejection reasons but does not alter the factor, invent a ridge, or implement fallback; BayesianPhysTwin remains responsible for exact caller-owned fallback.

A point-position Jacobian helper covers endpoint, marker, contact, and probe-position queries. Arbitrary nonlinear physical queries can supply their own local Jacobian and output metric.

Deterministic failure control

The checked evidence isolates one unobservable rotation-about-line direction under a complete identity prior and one frozen gate:

Case	Rank	Direct support	Variance reduction	Worst ratio	Decision
Point on observed line	6	1.000	0.909	0.091	admit
Distant off-axis probe	6	0.679	0.618	0.965	reject
Invalid full-rank completion, same probe	7	1.000	0.909	0.091	admit

The last arm is intentionally invalid. It demonstrates that query conditioning cannot repair a factor that has already fabricated information in its nullspace; preserving rank deficiency is necessary.

Files

• src/prob4d/query_observability.py: reports, gate, and point-query Jacobian;
• src/prob4d/query_observability_study.py: deterministic analytic control and CLI;
• evidence/query-observability-control-v1/result.json: checked mechanism evidence;
• tests/test_query_observability.py: eight focused tests, including prior-correlation and evidence-drift controls;
• docs/query-conditioned-observability.md: derivation, claim boundary, and prospective real-provider design.

Paper path

This supports a focused Prob4D companion-paper thesis:

Observability-aware probabilistic fusion of overlapping learned 4D predictions for guarded physical queries.

The material claim still requires one fresh real action-conditioned provider. The documented promotion path is PointWorld on garment-disjoint Flat'n'Fold groups, with source-only support/calibration, persistent sparse identities, frozen query metrics and thresholds, full-rank/ridge failure controls, grouped proper scores, accepted-update harm, worst-group regret, and exact fallback. Provider competence and downstream BayesianPhysTwin value remain separate endpoints. Causal4D is optional downstream evidence and cannot rescue an upstream failure.

Validation before push

The new module, study, and tests were syntax-checked locally. An isolated execution using the declared analytic factor passed all 8 focused tests and reproduced the checked JSON exactly.

Claim boundary

This PR establishes software and deterministic mechanism evidence only. It does not establish real-provider competence, held-out calibration, BayesianPhysTwin benefit, Causal4D benefit, deployment safety, or state of the art.

[FlorianPfaff]

Validation is complete on head 09128ad756b193a359239ec05006e663a615347c.

Green workflows:

• Fail-closed quality;
• provider batch preflight;
• cross-stack metamorphic invariants;
• security scanning; and
• complete Tests workflow, including the full suite on Python 3.10, 3.11, 3.12, 3.13, and 3.14, declared runtime-floor verification, wheel/sdist build, and installed-artifact CLI smoke tests.

The PR is now ready for scientific review. Its claim boundary remains deterministic mechanism/software evidence only; the PointWorld/Flat'n'Fold source qualification and later garment-disjoint held-out experiment remain the material empirical gate.