Re-ran the head-to-head on the current stack (multivon-eval 0.15.1 from
PyPI, DeepEval 4.0.2, RAGAS 0.4.3), gpt-4o-mini judge, temp 0, single run.

- RAGAS completes again. It errored on every case in the 0.9.8 harness;
  ragas 0.4.3 fixed that, so this is a real three-way comparison again
  (4/100 cases still error, ~13x slower than the others).
- ragtruth-sum n=100 headline on 0.15.1: multivon F1 0.729 (default 0.90),
  0.837 best-tuned (0.95); DeepEval 0.038 / 0.609; RAGAS 0.038 / 0.812.
  multivon shifted from 0.744->0.729 (within the ~+/-10pp CI at n=100;
  single-run, gpt-4o-mini temp-0 noise).
- Killed the stale "now 0.12.0 / refresh planned" snapshot note and the
  0.9.8 references; badges and tag bumped to 0.15.1 / 2026-06-26.

NOTE: run.py skips frameworks whose run0 file already exists, so the stale
multivon/deepeval raw files had to be deleted before this re-run actually
re-measured them (the first pass silently reused 0.9.8 outputs).