A hand-written tokenizer + recursive-descent parser that reads
the Lean source emitted by `toLeanSource` and reconstructs the
original meta-mirror value. Foundation/MetaParse.lean: 300
lines, faithful to the renderer's exact format.
Components:
· Token type (parens, ident chains, string literals, num literals).
· `tokenize : List Char → List Token` (partial; structural
decrease is implicit via helpers).
· `parseName?`, `parseClassifier?`, `parseMetaCTerm?`,
`parseArtifact?` — recursive-descent, return Option (T × tail).
· `MetaCTerm.fromLeanSource?` / `MetaClassifier.fromLeanSource?`
/ `MetaArtifact.fromLeanSource?` — top-level wrappers
demanding full input consumption.
Foundation/Meta.lean: derive `DecidableEq` on `MetaCTerm` (its
field types — Lean.Name, String, MetaClassifier — all have
DecidableEq). Switch FaceFormula.eq0/eq1 encoding from
`Name.appendAfter "_eq_0"` (string suffix) to a 2-component
`Name.mkStr (.mkSimple "eq0") i.name` form so reflection
round-trips by rfl with no string-suffix munging.
Foundation/MetaParse.lean: parsers are `partial def` because
the recursive calls land on output tails of helper readers,
which Lean can't see as structurally smaller without auxiliary
"consumes input" lemmas. Kernel-reducible round-trip is
deferred — `native_decide`-based tests in Infoductor/Test.lean
witness round-trip operationally for every meta-mirror arm.
Tests: 11 native_decide examples covering empty/ident/sym/app/
lam/comp/transp on MetaCTerm, always/meet on MetaClassifier,
empty/cterm on MetaArtifact (artifact uses rendering-equivalence
since Lean.Syntax in `.declAt` lacks DecidableEq).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
92fc4f9682