v0.1.4 adds attention_routing example (planning attractor measurement
via attention head analysis), CLT decoder vector steering mode in
steering_convergence, and comprehensive documentation comparing findings
with Anthropic's "Planning in Poems" attribution graphs work.

Key findings: L21:H5 is the dominant planning routing head in Gemma 2 2B,
planning attractor has a soft boundary (vs factual recall's hard boundary),
and suppress+inject produces 13x stronger routing than inject alone.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>