v0.9.1 — evidence wording polish

A patch release — no product behavior change. It makes the v0.9.0 evidence harder to misread.

• README badge/version updated: the stale hardcoded v0.8 badge is replaced with a dynamic GitHub-release badge (reads the latest release, so it can't drift again).
• "proof" softened to "evidence": the throwaway self-host demo line now reads "evidence that the mechanism can catch this kind of checked break on real code," preserving the caveat that it was a throwaway demo, not a committed artifact or benchmark figure.
• "5 limitations closed" clarified: v0.9.0 closed five binding-hardening issues, not all of dorian's strategic limitations. These remain open and visible: ambiguous symbols are skipped unless disambiguated; Python-only definition indexing; revalidate stays sidecar/store-driven; auto-binding is verify-only; trigger coverage is not behavior proof; real-world reproductions are scoped, not universal validation.
• Benchmark counts were re-checked against the machine outputs (summary JSON / generated doc / README) and agree (808 pairs · 63 domains · 122 artifacts · 122 claims · 408 mutations; checker_path recall 0.54 → bound 1.00; verdict precision 1.00, zero false BROKEN). The full benchmark was not rerun — this patch changed docs/version only — so the historical run provenance honestly stays dorian 0.9.0.
• Real-world scope clarified: a cited discovery catalog (12 public problems across 5 categories) was added to the protocol so the 5 report cases (3 hermetic reproductions + 2 documented) read as a selection, not a cherry-pick.
• Trigger-vs-truth framing preserved: binding widens the re-check trigger; the checker still decides truth; a watched file changing never makes a claim BROKEN by itself.
• Added a docs-polish guard test; full gate green (582 passed), ruff clean. The v0.9.0 tag was not modified.