ROADMAP v3: reorganize around transformerless LLM north star

Replaces v2's "what to ship next for the harmonic libraries"
framing with the 6-layer architecture from the new README:
substrate -> HBit -> JIT -> harmonic libs -> hybrid LLM
experiments -> infrastructure. Each layer's next-steps are
ranked by leverage toward the transformerless LLM north star.

Two priority items at the top:
- E1: find harmonic-attention primitives that beat softmax on
  a real task (THE blocker on transformerless)
- E2: find harmonic positional encoding that wins beyond L>=16
  (without distinct PE, no sequence model)

Three near-term high-priority items (move JIT from "works on
microbenches" to "works on the harmonic libraries"):
- J1: dict support in codegen (2-3 sessions)
- J3: fallback-to-tree-walk for one builtin within JIT'd fn
- J4: float-typed Div/comparison in bytecode compiler
- L1: rewrite harmonic_anomaly to use array-of-hashed-int
  (alternative to J1 — half a session of library refactor)

Deprecated v2 items framed around "OMC as ML library
standalone" or "competing with sklearn generally" — superseded
by the LLM north star. The harmonic libraries remain valuable
as proof points but they're not the project's purpose.

Five v3 principles + five anti-goals make the framing
load-bearing for future planning sessions.

Includes today's session summary: 14 commits across Sessions
A-H + Path A-D + README reframe; 41 codegen tests passing,
149 core unit tests, 18 OMC harmonic-lib tests, engine parity
44/45 byte-identical.

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

Author: RandomCoder-lab