LLM discoverability: builtin metadata registry + introspection + did_you_mean

The discoverability layer that lets LLMs write OMC code without seeing
it in training data. Adds a structured BuiltinDoc registry in
omnimcode-core/src/docs.rs — currently 97 entries across 14 categories,
13 of them flagged unique_to_omc for the substrate-aware primitives
that have no Python equivalent.

New runtime introspection builtins:
  omc_help(name)             — dict of {name, category, signature,
                                description, example, unique_to_omc}.
                                On miss, returns {found:0, did_you_mean:[...]}.
  omc_list_builtins(cat?)    — all names, optionally filtered by category
  omc_categories()           — every distinct category
  omc_did_you_mean(name, k?) — closest-name suggestions by edit distance
  omc_unique_builtins()      — only the OMC-only stuff (substrate, autograd,
                                lazy generators, harmonic ops)

Error messages: "Undefined function: arr_softmx" now becomes
"Undefined function: arr_softmx (did you mean: arr_softmax?)" — by
running the unknown name through the same Levenshtein suggester.

New CLI mode: `omc --gen-docs > OMC_REFERENCE.md` streams the full
registry as Markdown. The committed OMC_REFERENCE.md (1055 lines)
is the canonical prompt-injection material for LLM clients — a few
KB of structured signatures + descriptions + examples that an LLM
can read once and write idiomatic OMC.

Tests: 13 cases covering omc_help on known/unknown/substrate/standard
builtins, category filtering, did_you_mean close/no-match/limit
semantics, the unique-builtins surface, and that unknown-function
runtime errors carry the suggestion in their message.

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

Author: RandomCoder-lab