refactor(transparency): resolve method families via own algorithm_registry; drop framework hardcoding

[stanlrt]

Summary

Removes bootstrap-era hardcoded ShapExplainer/CaptumExplainer knowledge from the framework-agnostic transparency layer (issue #318), so third-party plugin adapters are handled correctly instead of silently mishandled.

Investigation showed the fix is a pure simplification, not a new abstraction: every real explainer — in-tree or plugin — is a decorated adapter carrying its own algorithm_registry ClassVar (mandated by @adapters.transparency), so it always resolves via the own-registry lookup. The framework-label path and the cross-framework algorithm-string scan were dead in production (only test stubs reached them), and the factory's "shap"/"captum" name-sniffing gate actively skipped the visualiser semantic-compat check for plugins named neither.

Changes:

• transparency/semantics.py: method_families_for_explainer / _hints_for_explainer now resolve solely via type(explainer).algorithm_registry (same shape as robustness/semantics.py). Deleted the dead helpers _adapter_class_for_framework, _explainer_framework, _normalise_framework, _method_families_for_algorithm, and the unused AbstractSet import.
• transparency/factory.py: deleted _requires_registry_semantics (the name-sniff gate) and the two hardcoded explainer imports; check_explainer_visualiser_semantic_compat now always runs and dropped its unused explainer_target param.
• Tests: stubs rewritten to use real adapter instances / real algorithm_registry; deleted tests that only exercised the removed fallback paths.

Net: 25 insertions / 225 deletions. No behavior change for shap/captum; the only intended behavior change is that the visualiser semantic-compat check now also runs for adapters named neither (i.e. future plugins).

API note (reviewers): removed the algorithm= keyword param from the public infer_output_space and the explainer_target param from check_explainer_visualiser_semantic_compat. Both were dead — no in-tree production caller passed them (only tests did). Pre-1.0, treated as a clean removal, not a breaking change.

Closes #318.

Checklist

• CI — Required checks are green
• Breaking changes — Not a breaking change: the only removed surface (infer_output_space(algorithm=...), check_explainer_visualiser_semantic_compat(explainer_target=...)) had no production callers. No ! in the title.
• Contributor guide — I’ve read Pull requests and commit messages before requesting review.

Optional

• Issue (optional) — Closes Transparency semantics/factory hardcode SHAP/Captum adapter classes (silently breaks plugin adapters) #318.
• Docs (optional) — No doc change needed; no live user/contributor doc described the removed surface (verified).
• Tests (optional) — Test stubs migrated to the own-registry path; dead-path tests removed.

[stanlrt]

Folded in two de-hardcode follow-ups surfaced by a parallel dead-code audit (same #318 smell family), commit 9cc2a6ae:

• utils/diagnostics.py: replaced the hardcoded ("raitap.transparency", "raitap.robustness") provider tuple with a register_configs() call — now aggregates third-party libs across all families + discovered plugins (drift-proof, plugin-inclusive), instead of a hand-maintained 2-family subset.
• robustness/_ci_router.py: derive the ci_method value set from get_args(CIMethod) instead of re-declaring the ("wilson", "clopper_pearson") literal (no-duplicated-Literals rule). Error wording unchanged.

Both verified locally (test_diagnostics 16, test_ci 6 green).