v0.9.0

Make the model's own claims trustworthy, then start catching the defects a green model used to hide.

Highlights

• Source trust & extraction completeness — rank sources by authority (contradictions resolve by weight, not reading order), record two that disagree as a first-class conflict, stage behavior candidates before committing them, and measure the share of each source that actually reached the model (get_source_coverage).
• Model what actually happens — first-class action outcomes (declined, timed out, partial), external dependencies + the invoke_operation effect with their timeout/retry/idempotency contract, event delivery guarantees (a required handler failure fails the emitter; transactional rolls it back), and scoped invariant relaxation for repair actions.
• Sharper verification — bounded model checking now explores parameter domains (not just defaults), an honest confidence matrix in score_feature (the weakest dimension is the headline), decision-table analysis in get_spec_gaps, and an evidence-first unspa check --strict gate.
• Feature-level prose acceptance criteria — capture Given/When/Then edge cases as documentation, the honest complement to model-checked scenarios: rendered, searchable, carried in the model, but never simulated or scored.
• Core features — declare a project's product pillars once and group features under them precisely: a reserved core: tag backed by a project-level registry, so you filter by core feature instead of by a sea of tags.

Full detail in CHANGELOG.md.