Summary

• The lexer built its symbol character class only from the glyphs of registered symbols, so any other non-ASCII glyph matched no token and String#scan silently skipped it. Unknown ASCII words still matched the baseline [A-Za-z0-9_]+ class and reached validation, but unknown glyphs vanished:
  • Unit(1, 'λ') returned a unitless 1
  • Unit(1, 'm λ') silently became 1 m
  • Unit(1, 'm/λ') raised a misleading SyntaxError: Unexpected token /
• Now the lexer tokenizes any run of non-operator, non-whitespace characters as a symbol, so every unrecognized glyph survives lexing and fails loudly as Undefined unit <symbol> through the same validation path unknown ASCII symbols already used — and the error names the real culprit (Undefined unit λ) rather than a neighboring operator.
• As a side effect the tokenizer no longer depends on what is registered, so it is a plain constant again instead of a per-load rebuild. Registered glyphs (Ω, °C, µ, %, units loaded at runtime) still parse, and to_s round-trips are preserved.

Breaking: parsing an unrecognized unit symbol now always raises TypeError: Undefined unit … instead of sometimes silently yielding a dimensionless value. This extends the contract that unknown ASCII symbols already followed to unknown glyphs of any alphabet.

Test plan

• bundle exec rake — both phases green (140 no-DSL + 142 DSL examples, 0 failures)
• New specs cover: unknown ASCII symbol; unregistered glyph alone; glyph implicitly multiplied with a known unit; glyph adjacent to an operator; and a unit whose defining system has not been loaded yet (raises until loaded, then resolves)