feat(core): native addressed-memory primitive (amem_*) + write-don't-train dot-connector research arc

Core language feature (shippable):
- omnimcode-core: new src/addressed_memory.rs — in-core kNN datastore + k-means IVF index
  (thread-local handle registry). Five OMC builtins amem_new / amem_write / amem_index /
  amem_search / amem_len, wired through interpreter dispatch + docs/omc_help. 2 Rust unit +
  4 OMC-level tests; full suite no-regression. Sits alongside haddr/cas/heap as the
  memory+index substrate.

Research arc (experiments/transformerless_lm) — agnostic, pre-registered/A-B-validated; see
GENERATOR_PLAN.md (BUILD-1..21) and MANIFEST.md for the navigable map:
- write-don't-train addressed memory: +11.35% on an unseen domain with zero gradient (AM-5);
  IVF wins the index frontier over substrate geometry (159-run study); code is +48% vs prose
  +14% injectable.
- BPE tokenizer, chunked span-copy, novelty pressure; generator true-ceiling study
  (substrate ~7x param-efficiency; the permanent edge is structural/live knowledge).
- dot-connector stack: grounded multi-hop navigation (per-hop recall 1.00 via addressed
  inverted index), self-learned meaning-judge (held-out relatedness AUC 0.83 vs lexical 0.19),
  the closed loop (meaning-guided nav 3.2x more coherent than blind BFS), calibrated narrator
  voice, relation-shape analogy/joke mechanism; cross-domain analogy real across 5 domains
  (shape-cosine 0.774 vs 0.406 shuffled).
- char-substrate skills (tokenization-blind tasks); chat router.
- law-clean throughout: no hand-coded dictionaries (verbs derived by inflection paradigm,
  knowledge read from a corpus file).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Author: RandomCoder-lab