feat(framework): [B] strengthen diagnostic evidence model

Reframes the framework from privileged causal explanation to disciplined
hypothesis generation and testing. Key changes:

- scoring: replace overall_confidence float with DiagnosticSupportProfile
  (per-hypothesis scores; multi-causal and unknown outcomes; compat property)
- coherence: rename genuine/performed/mixed to high-continuity/partial/fragmented
  to describe claim-continuity patterns without implying truth claims
- crosscheck: add pair_provenance to ABTestResult; add run_ab_test_with_user_task
  for analyst-authored A/B pairs that avoid machine-inversion artifacts
- report/cli: expose support_profile in structured output; add pair_provenance;
  retain backward-compat color maps for old report values
- framework docs: update Rule 4 (ratchet measures continuity, not honesty),
  Rule 5 (baseline distinguishes user expectation / system contract / policy),
  falsifiability evidence hierarchy, taxonomy with unknown/multi-causal rows
  and discriminating-probe table, manual-diagnosis workflow
- prompts: update R4 text and 3.2 card with pair-provenance documentation;
  keep both prompts.yaml mirrors byte-for-byte identical
- calibration: add compat label map so legacy reference set stays intact
- docs/migration.md: document field renames and new APIs
- validation/next-stage-benchmark.md: specify credible next-stage study design

Quality gates: ruff clean, pytest 382 passed 78% coverage, python -m build ok.

Assisted-by: Claude:Sonnet-4.6
Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>