README: reframe around the transformerless LLM goal

OMC isn't a wrong-trajectory project. The recent work (substrate
fix to log_phi_pi_fibonacci, HBit dual-band wired through the JIT,
harmony-gated branch elision, harmonic libraries on real data,
LLVM JIT pipeline) is all building toward what the architecture
was always meant to be: a substrate capable of hosting a
transformerless LLM.

The previous README sold OMC as four separate things — anomaly
detector, Python bridge, package manager, self-hosting language.
True but disconnected. The new README presents OMC as one thing:
a 6-layer stack (substrate → HBit → JIT → harmonic libs → hybrid
LLM experiments → infrastructure), with the transformerless LLM
as the architectural north star at the top.

What's framed differently:
- The 10 hybrid_llm experiments are now a top-section feature,
  with the headline finding (HBit cross-cutting tension AUROC 1.0
  as reference-free OOD) called out alongside the negative
  findings (OmniWeight loses to softmax on perturbed query;
  multi-channel harmonic PE loses to sinusoidal at L≥16).
- The harmonic_anomaly library numbers are still here but
  framed as evidence the substrate is useful for ML, not as
  the product itself.
- The JIT (Sessions A-H + Path A.1-A.4 + Path D) is now first-
  class in the layer stack, with honest links to bench results
  and the gap measured in docs/jit_real_world.md.
- Status section is explicit about what's NOT shipped: the
  transformerless LLM itself, per-component harmonic wins on
  PE/attention, dict/string codegen for harmonic libs to JIT,
  AVX-512 widening, float-typed Div.
- Repo layout table replaces the looser "what's in the box"
  enumeration so newcomers can navigate.

The thesis statement: harmonic primitives win as detectors today
(experiments validated), need to win on primary computation
paths to enable transformerless. The substrate work shipped so
far is what makes that future work possible.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Author: RandomCoder-lab