Legacy model line: two .exp files with the same file name in different directories collapse to one key and the first is silently dropped

The bug is in pybnf/config.py lines 2838-2839 and 2855 where _file_prefix strips the directory, causing collisions. Start by reading the _load_exp_data and _file_prefix methods. Check how _data_map is populated at lines 1022 and 1515. Write a test that loads two .exp files with the same basename from different directories and verify that an error is raised, matching the behavior of _load_analytical_data. Run existing tests to ensure no regression.

_load_exp_data stores each data file at ed[m][self._file_prefix(ef)] (2855). _file_prefix keeps only the basename without extension (2838-2839). Two files in _data_map[m] (filled at 1022 and 1515) with the same basename therefore write the same key, and the last one wins with no warning. _check_actions (2871-2886) builds a set of prefixes, so it also reports nothing. The analytical-objective loader (_load_analytical_data, 1466-1472) raises a clear error on the same stem collision, so a collision is clearly meant to be an error.

Failure scenario

'model = parabola.bngl : rep1/par1.exp, rep2/par1.exp' (replicate files kept in two folders). Config loads without a message, exp_data has one 'par1' entry holding rep2's values, and the fit scores only half the data: the reported n, objective and information criteria silently omit rep1.

Reproduction (independent re-run)

Setup: rep1/par1.exp is the demo data (true v=(0.5,1,3)); rep2/par1.exp is the same data with y+2. Configs were loaded with load_config (pybnf from pybnf/init.py).

Load: both.conf gives exp_data keys=['par1'], y(t=0)=[45.0] (rep2 only), 21 data rows scored, and mapping={'parabola': {'par1'}}. No warning was raised.

Objective from PyBNF's real evaluate_multiple on both.conf, compared with the correct value (rep1-only conf plus rep2-only conf, which matches a hand numpy value of 0.5*chi2 over both files):

• v3=4 (correct pooled optimum): PyBNF 10.5, correct 21.0.
• v3=5: PyBNF 0.0, correct 42.0. PyBNF calls v3=5 a perfect fit when it is not the optimum.

End-to-end CLI fit (python -m pybnf -c both.conf -o, DE with 20x30 budget, seed 1): best fit v3=5.048 with obj=0.0786. The correct pooled optimum is v3=4. The console printed "AIC=83.3481 BIC=88.5611 (k=3, n=42)": n=42 is one file's 21 rows x 2 columns, where the correct n is 84. No warning appeared anywhere in the log.

Reachability

model = parabola.bngl : rep1/par1.exp, rep2/par1.exp, with fit_type = de and objfunc = chi_sq, run through python -m pybnf -c both.conf -o. It also applies to legacy mutant ... : dirA/xyz.exp, dirB/xyz.exp lines, which extend the same _data_map at config.py:1515.

Where

pybnf/config.py:2855

Found in a whole-codebase audit for silently wrong results (2026-09-23); the reproduction above was re-run independently of the original finding.