transformerless_lm: v3 results — phi-power ties fib at the saturation floor

Three tier bases now tested (per_row scale, n=16):
  fibonacci/nor       dense_crt +0.74    tied_substrate +0.59
  fibonacci/rec       dense_crt +0.77    tied_substrate +0.59
  phi_power           dense_crt +0.77    tied_substrate +0.59

phi_power (continuous Binet limit, ratio exactly phi) TIES the best
Fibonacci configuration but does not beat it. All three bases hit
the same floor:
  dense_crt:       ~+0.74 nats (~0.30 per-row scale gain over per_tensor)
  tied_substrate:  ~+0.59 nats (consistent 0.15 nat improvement over dense_crt)

The substrate basis is not the bottleneck. The bottleneck is that
weights trained with unconstrained float gradients have arbitrary fine
structure that no log-spaced tier set can recover post-hoc. Standard
4-bit linear quantization on language models loses <0.1 nats; we are
6-7x worse.

What this tells us:
  - Substrate as a POST-HOC compression scheme is bounded by how
    Fibonacci-shaped the float weights happen to be at convergence.
  - tied_substrate weights are MEASURABLY more substrate-shaped than
    dense_crt weights (+0.15 nat advantage at quantization), which
    confirms Principle A x Principle B compose, just not enough.
  - To clear the 0.1 nat bar we need quantization-aware training:
    weights snapped to the tier set inside the gradient loop, with
    straight-through-estimator surrogates so gradients can flow.

This closes the "fluid Fibonacci" question: making Fibonacci continuous
(phi_power) doesn't help because the problem is on the weight side,
not the tier-basis side.

Author: claude