[pederbe]

Multipart parts without a Content-Type currently pass through text decoding before reaching TempFile or byte-slice form guards. This replaces invalid UTF-8 and can strip or transcode byte-order marks, changing uploaded files even when a filename is supplied.

Retain the original part bytes and add defaulted raw-value callbacks to FromForm and FromFormField. Derived forms and built-in wrappers forward those bytes to binary guards, while existing text guards keep their decoding and value/data dispatch behavior. Headerless parts retain their existing buffered storage and data-form limit; explicit Content-Type parts keep streaming through the data-field path.

Custom form wrappers that forward value fields must also forward push_value_with_bytes to preserve binary contents in their inner guards. Transformed text retains both the original bytes and the decoded text. Regression coverage includes every byte value, both filename variants, byte-order marks, nested forms, legacy callbacks, limits, and cleanup.

With the dependency versions resolved locally, the full workspace build encounters a database dependency conflict that also reproduces on the unchanged base. Core tests and the examples suite pass.

Fixes #2934.

[SergioBenitez]

This is way over-scoped. The fix for the issue is way simpler. This is also clearly AI without little regard for human oversight.