The analysis pass between a front-end (Aether.Parser, or one of the ABNF/BNF/EBNF/PEG importers) and either backend (Grammar.VM or native codegen):

• reference checks -- every RuleRef must name a declared token or rule; checked first, since the other passes assume a valid namespace.
• left-recursion rewrite -- a rule that references itself as its own leftmost derivation is rewritten from A := A op b | base into the iterative A := base (op b)*, the standard PEG/packrat transform. Only direct self-recursion through a bare (uncaptured) leading reference is rewritten automatically; anything else (indirect cycles through other rules, or a self-reference buried under a capture) is reported as an error instead of guessed at.
• possibly-empty-match repetition detection -- Star/Plus wrapping an expression that can match the empty string never terminates; flagged as an error, not just a lint, since it's a real runtime hang, not a style nit.
• duplicate-alternative lint -- two structurally identical alternatives in the same Choice are dead code under PEG's first-match-wins semantics -- the second can never be reached.

Predefined-token override-ordering validation is already enforced by Aether.Parser itself, as it happens -- it's entirely about Aether's own five predefined tokens (DIGIT/ALPHA/ALNUM/SPACE/HEX), so it doesn't generalize to other front-ends the way the checks above do, and isn't repeated here.