Stop write_gsf mutating the caller's DataFrame

[lispandfound]

Fixes #89

Problem

write_gsf filled its optional columns by assigning straight into the caller's DataFrame:

# before
if "init_time" not in gsf_df:
    gsf_df["init_time"] = -1
if "slip" not in gsf_df:
    gsf_df["slip"] = -1
if "loc_rake" not in gsf_df:
    raise ValueError("The DataFrame must have a 'loc_rake' column.")

A frame handed to write_gsf came back carrying two columns of -1 sentinels it never had. Since write_gsf is a serialisation function, nothing in its signature or docstring suggests the input is consumed.

Because the assignments sat above the loc_rake check, the mutation also happened on the failure path — write_gsf could raise, write no file, and still leave the caller's frame modified.

Impact

df = pd.DataFrame({... "loc_rake": [0.0], "seg_no": [0]})
write_gsf(df, path)
list(df.columns)
# before: [..., seg_no, init_time, slip]   values -1
# after:  [..., seg_no]

Sentinel -1 values silently persisting in a caller's slip/init_time columns are the kind of thing that reads as real data downstream.

Changes

• source_modelling/gsf.py — fill defaults on a shallow copy (copy(deep=False); only new columns are added, the existing column data is never written through, so a deep copy would be wasted work on large fault meshes). Move the loc_rake validation above the defaulting so the failure path leaves nothing behind.
• tests/test_gsf.py — add test_write_gsf_does_not_mutate_input, covering both the success and the raising path.

No change to the file written — the same columns are serialised with the same defaults.

Verification

• New test fails without the fix, passes with it.
• pytest tests/test_gsf.py — 13 passed.
• ruff check, ruff format, numpydoc lint clean.

🤖 Generated with Claude Code