vindex 0.5.0 — the token mixer, first-class

The architecture-honesty release, driven by a hybrid hero model:

• describe layer.N.mixer — the token mixer as a semantic component: its programme (GATED DELTANET / GATED ATTENTION / SOFTMAX ATTENTION, from the operation plan) and every operand with its plan-assigned role — fused recurrent q|k|v, decay, write-strength, output gate, conv, log decay, timestep bias, norm, output.
• vindex layers — each layer's mixer programme and FFN kind, ranges collapsed: three seconds that explain why a hybrid model is an interesting quantization subject.
• precision --matrix is programme-aware — layers group by mixer, each group carrying only the columns its programme computes with, plus MODEL SURFACES (embedding, head, towers) at their own derived bits. No architecture forced into another's schema.
• The attention.q refusal is corrected: DeltaNet layers do have recurrent q/k/v — the message now says what the mixer is and points at layer.N.mixer.

vindex update from any earlier release installs this one.

What's Changed

• test(vindex-cli): DeltaNet refusal proven on an encoded hybrid by @chrishayuk in #334
• feat(vindex-cli): the token mixer becomes a first-class component by @chrishayuk in #335

Full Changelog: vindex-v0.4.0...vindex-v0.5.0