[edkerk]

Stacked on #3 — review that one first; this diff is only the new data layer.

What changed

• src/overflow/proteomics.py — reading, the replicate quality filter,
  mmol/gDW → mg/gDW conversion, the f-factor and the protein-content
  rescaling.
• data/uniprot.tsv — vendored molecular masses (GECKO develop4,
  same commit as the models), covering all 2928 measured proteins.
• tests/test_proteomics.py — 31 tests.

The filter is checked against MATLAB's own output, not just described

filter_ProtData reports mean + 1.96·SD, not the replicate mean — easy
to get wrong, and invisible afterwards. The legacy
results/modelGeneration/modifiedEnzymes_*.txt files record the abundances
the MATLAB run carried forward, so they pin the filter exactly:

• CN4: all 7 entries match to 1e-12 relative. CN4 flexibilized nothing
  before integration, so every entry is raw filter output.
• The other conditions match exactly where the minimum-usage pass did not
  intervene, and are strictly higher in the legacy file where it did — the
  test asserts that direction, so a filter that drifted downward would fail.

Two data issues found

1. P61830 is on two rows (the duplicated histone, one row per gene) with
   identical values. The reader collapses it and raises if duplicated rows ever
   disagree, rather than letting pandas return a DataFrame where a Series was
   expected.
2. A missing replicate drops the protein. MATLAB's std propagates NaN, so
   such a row can never pass the variability test — the 2/3 presence check
   above it only ever fires on explicit zeros. Substituting 0 for NaN, which is
   the natural Python reading, would keep those proteins with a downward-biased
   abundance. Ported faithfully and tested.

Filter output

Condition	kept	removed	measured mass [mg/gDW]	Ptot [mg/gDW]	Ptot rescaled
CN4	2107	28.0%	418	500	0.5725
CN22	2116	27.7%	429	550	0.5890
CN38	2190	25.2%	269	330	0.3524
CN75	2147	26.7%	220	270	0.2958
hGR	2004	31.6%	294	410	0.4448

Measured mass lands at 72–86% of the stated protein content, which is what a
correct unit conversion should look like. A test brackets this per condition,
so a unit slip fails loudly rather than propagating into enzyme constraints.

Deliberate fidelity choices

• The rescaling Ptot * sumPfilt/sumP is a molar ratio in the original, not
  a mass ratio, and is ported as such. Because the filter reports
  mean + 1.96·SD, the kept fraction exceeds one and the protein content
  scales upwards — visible in the table above. Preserved rather than
  "corrected".
• The f-factor is taken from the unfiltered means, as in the original, which
  covers more proteins than the filtered set. A test asserts it agrees with
  geckopy's calculate_f_factor to 1e-12.