Normalize Lean 4.33 docstrings before pydantic validates them

[Gabrielebattimelli]

Why

Indexing is still blocked. jixia now builds fine against Lean 4.33 (the patch loop works), but the smoke test rejects its output — every declaration with a docstring fails:

0.modifiers.docString
  Input should be a valid tuple [type=tuple_type, input_value='The charge density. ', input_type=str]

Lean 4.32 emitted docString as a [text, bool] pair; 4.33 emits a bare string.

Why the previous fix didn't work

#26 widened Modifiers' annotation and called model_rebuild(force=True). That fixes Modifiers.model_validate(...) in isolation — which is presumably how it was checked — but not the real path.

Declaration and Symbol compile their own nested copy of Modifiers' schema when their classes are created. Rebuilding Modifiers afterwards leaves those copies validating against the original tuple-only type. Confirmed directly:

Modifiers.model_validate(...)          -> OK
Declaration.from_json_file(...)        -> docString tuple_type error

I also tried clearing __pydantic_core_schema__ / __pydantic_validator__ on the dependents and rebuilding — the error persisted (pydantic 2.13.4).

What

Normalize the raw JSON to the pair form before pydantic sees it. Every read goes through RootModel.from_obj, so this covers Declaration, Symbol, and anything else jixia loads, without depending on pydantic schema internals.

Verified

Using the shim exactly as written in the file:

4.33 string  -> ('doc', False)
4.32 tuple   -> ('doc', False)
none         -> None

Both Lean versions parse, so this doesn't break if a patch selects a 4.32-shaped build.