parse_reference skips rows with a non-numeric share, but no test shows it or compares it with the web engine #1002

Where: faircode/profiler.py parse_reference() (the except (TypeError, ValueError): continue, lines 541-542), assets/profiler-engine.js parseReference(), tests/test_profiler.py, tests/test_js_parity_profile_thresholds.py.

Current gap: a reference row whose share cannot be read (n/a, an empty cell, ?) is dropped without a message, and the remaining shares are used as they are. Coverage shows the branch never runs. Neither the silent skip nor the question of whether the web engine drops the same rows (and keeps the same per-column scale decision) is tested, so a baseline with one bad cell could change which columns count as percentages in one engine only.

Repro:

>>> import pandas as pd; from faircode.profiler import parse_reference
>>> parse_reference(pd.DataFrame({"column": ["sex"] * 3, "group": ["M", "F", "X"], "share": [0.5, "n/a", 0.5]}))
{'sex': {'M': 0.5, 'X': 0.5}}      # the F row vanished with no note

Suggested approach: add the case to the Python tests and to the existing parseReference parity test, and decide whether a skipped row should add a note to the result (the CLI already prints notes for similar drops).