catalog growth: errors 702 → 809; token dict +145 idiom entries

Tokenizer: 145 more entries covering common OMC code shapes —
type-check patterns (if type_of(), is_attractor(), is_defined()),
dict ops, accumulator initializers (h sum = 0;, h xs = [];,
arr_push(out, ), matrix/ML variable names (h W = , h grad = ,
h loss = ), autograd phrases (tape_var(0.0, h L = tape_sum(),
LLM-introspection calls (omc_help(", omc_explain_error(),
common test phrasing, control-flow continuations, error-message
fragments. Compression on test files and tutorials goes up.

Errors: 107 more patterns covering runtime + types + integration:
  math: overflow/underflow/modulo-by-zero/log-of-nonpositive/
        sqrt-of-negative/asin-out-of-domain/pow-zero-zero
  runtime: infinite loop, function redefinition, variable
        shadowing, class not found, method not found,
        inheritance cycle, safe applied wrong, self required,
        uninitialized var, write to immutable, ambiguous overload
  generators: yield outside generator
  imports: cycle / not found
  iteration: not iterable, range OOB, nil pointer
  match: not exhaustive, type tag mismatch
  regex: too complex, invalid escape
  memory: OOM
  python interop: null, type error, not initialized, module
        missing, attribute missing
  autograd: tape not reset, gradient zero (disconnected),
        matmul wrong rank
  io: disk full, file too large
  hashing: hash mismatch
  tokenizer: invalid token ID, empty stream
  code_intel: circular alias

Plus 81 per-builtin variants: "<name>: not an array" and
"<name>: not a number" for the most-used builtins, and "<name>:
parse error" for every code_intel input function.

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

Author: RandomCoder-lab