Fold builtin calls on the authored literal, not the coerced one

[e54-bot]

Parent: #366. Found by the builtin probe in #370 (gap fold-after-coercion).

Goal

Make builtin folds see the literal the author wrote, as the pinned OverPy 9.7.10 does, instead of the parameter's canonical form.

Context

Lowering rewrites False, True, 0 and an empty array to the parameter's canonical form (normalize_value_with_coercions in compiler/lowering.rs) before optimization. A fold on the rewritten value then differs from the reference, which folds on what was written: min(false, 3) folds to 0 natively but stays Min(False, 3); [1, 2].index(0) keeps Null natively but folds to -1; hudText(..., [], ...) becomes Null; .replace, .slice behave alike. 38 probes (6 functions) differ; after workshop-rs 0.6.4 parsing treats the coercible spellings as one structure, so only the folds show.

Scope

• Apply the coercions after optimization, or stop applying them where the validator now accepts the authored form, without changing the structure of currently converged output.
• Oracle-backed fixtures for the folds above.

Non-goals

• Changing workshop-rs.

Acceptance criteria

• The fold-after-coercion gap is removed from tools/overpy/probe-gaps.json.
• Existing integration_tests and the corpus stay green (12 semantic-equivalence tests depend on today's normalization).