pydata egress: crashes on DataFrame columns with non-identifier names

[discreteds]

Problem

pydata.egress converts DataFrame columns to namedtuple field names during the to_list_of_dataclasses() path. If a column name isn't a valid Python identifier (e.g., foo-bar, class, or contains spaces/special chars), the namedtuple construction fails.

Reproduction

import polars as pl
from mountainash.pydata import PydataEgress

df = pl.DataFrame({"valid_col": [1], "foo-bar": [2], "class": [3]})
# Attempt to convert to dataclasses or named tuples → crash

Context

Discovered while designing mountainash-wearables' data querying layer. When conforming API responses with keep_only_mapped=False, unmapped provider columns (which can have arbitrary names like average_heartrate_bpm, foo-bar, or even reserved words) flow into egress and crash.

Workaround: Use keep_only_mapped=True in TypeSpecs to avoid the issue entirely. Raw data is preserved via a sidecar list, not through unmapped columns.

Suggestion

Egress should either:

1. Sanitize column names before namedtuple construction (replace invalid chars with _, prefix digits)
2. Skip/warn on non-identifier columns rather than crashing
3. Use a dict-based intermediate instead of namedtuple when column names aren't all valid identifiers

This is low priority since keep_only_mapped=True avoids it, but it's a surprising failure mode for users who don't expect column names to matter.