Web profiler's JS engine silently ignores an unmatched reference baseline, unlike the CLI

[yakew7]

Where: assets/profiler-engine.js's profile() (the reference-handling branch) vs faircode/profiler.py:412-424.

The gap: faircode/profiler.py explicitly raises when none of a --reference file's columns match a profiled dimension (if not any(d["name"] in o["reference"] for d in dimensions): raise ValueError(...)). The JS engine has no equivalent check - it applies the reference silently whenever o.reference is truthy, with no validation.

Repro (verified against both engines with the same mismatched reference):

# Python
>>> profile(df, {}, {"reference": {"totally_wrong_col": {"a": 0.5, "b": 0.5}}})
ValueError: reference file's column(s) don't match any profiled dimension: totally_wrong_col

// JS - same input, no error, reference silently ignored
const r = E.profile(table, {}, {reference: {totally_wrong_col: {a:0.5,b:0.5}}});
r.dimensions[0].reference  // undefined - silently did nothing

Why it matters - this is reachable through the live web UI: assets/profiler-ui.js's reference-upload handler lets a user upload any CSV as a reference baseline, and on a successful read it unconditionally sets referenceStatus.textContent = '⚖ scored vs ' + f.name regardless of whether the reference actually matched anything. A user who uploads a reference file with a typo'd or wrong column name sees a confident "scored vs myfile.csv" status while the reference silently did nothing - the same silent-no-op class of bug already fixed for the CLI (closing #364).

Suggested fix: port the same validation into the JS engine's profile(), and have profiler-ui.js call showError() when it throws, instead of showing the reference as successfully applied.