Fix flaky Weber_BMC2015 gradient check with AMICI-native RobustConsistency

[dweindl]

Summary

• Fixes the long-standing intermittent test_benchmark_gradient[Weber_BMC2015-*-unscaled] failures (Check flaky test_benchmark_gradient / Weber_BMC2015 #3078): fiddy's Consistency success checker only verifies that forward/backward/central methods agree with each other at a given step size ("self-consistent"), then blends every self-consistent size into the final value. A step size can become small enough that all methods sample points within the target function's floating-point noise floor and become spuriously self-consistent while biased away from the true derivative -- this hit parameter a32, which is several orders of magnitude smaller than Weber_BMC2015's other free parameters.
• Adds RobustConsistency to amici.adapters.fiddy: a subclass of fiddy's Consistency that additionally rejects step sizes that are self-consistent but inconsistent with the majority of other step sizes, via iterative, order-independent, MAD-based outlier rejection.
• Switches every remaining Consistency usage in AMICI to RobustConsistency, in both tests/benchmark_models/test_petab_benchmark.py and python/tests/adapters/test_fiddy.py -- there's no reason to keep using the plain checker anywhere now, since RobustConsistency only depends on fiddy's released Consistency base class.
• Adds unit test coverage for the new class in python/tests/adapters/test_fiddy.py (12 tests, including a regression test reproducing the flaky-check mechanism from a minimal synthetic function).
• Unskips test_benchmark_gradient's unscaled (linear-scale) parametrization for 5 benchmark models -- Elowitz_Nature2000, Borghans_BiophysChem1997, Sneyd_PNAS2002, Bertozzi_PNAS2020, Zheng_PNAS2012 -- that were previously skipped because their finite-difference gradient checks only failed on linear scale, the same failure mode RobustConsistency guards against (step sizes spuriously self-consistent but biased, more likely when parameter magnitudes vary widely on linear scale). Verified locally that all 5 now pass reliably (forward and adjoint). Brannmark_JBC2010 remains skipped there -- it still fails independently of this fix. test_nominal_parameters_llh_v2's skip list is left untouched.

Background

This was originally proposed upstream as ICB-DCM/fiddy#77. After review feedback (the fiddy maintainer preferred not to add a parameter-magnitude-calibrating statistic to a general-purpose library, and suggested per-parameter/relative-step-size tuning instead), that PR was closed in favor of implementing the fix directly in AMICI. This PR supersedes the earlier approach of pinning CI to an unmerged fiddy branch -- that pin has been reverted, since RobustConsistency now lives in AMICI and only depends on fiddy's released Consistency base class.

Test plan

• pytest python/tests/adapters/test_fiddy.py -v -- 15/15 passed (3 pre-existing + 12 new RobustConsistency/_reject_outliers tests)
• pytest tests/benchmark_models/test_petab_benchmark.py -k "Weber_BMC2015 and unscaled" -- previously-flaky cases now pass reliably
• Spot-checked other benchmark models (Boehm_JProteomeRes2014, Fujita_SciSignal2010, both scales/sensitivity methods) for regressions -- all pass
• Verified the 5 newly-unskipped models pass their unscaled gradient checks (forward and adjoint) locally
• Benchmark Collection CPP GHA job passes on this branch

🤖 Generated with Claude Code

[codecov]

Codecov Report

❌ Patch coverage is 95.08197% with 3 lines in your changes missing coverage. Please review.
✅ Project coverage is 78.02%. Comparing base (3aebbe8) to head (dbfa0e6).
⚠️ Report is 3 commits behind head on main.

Files with missing lines	Patch %	Lines
python/sdist/amici/adapters/fiddy.py	95.08%	3 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #3227      +/-   ##
==========================================
- Coverage   78.62%   78.02%   -0.60%     
==========================================
  Files         318      318              
  Lines       21102    21167      +65     
  Branches     1487     1487              
==========================================
- Hits        16591    16516      -75     
- Misses       4503     4643     +140     
  Partials        8        8              

Flag	Coverage Δ
cpp	72.06% <95.08%> (+0.04%)	⬆️
cpp_python	36.45% <0.00%> (-0.12%)	⬇️
petab	48.13% <0.00%> (-0.17%)	⬇️
petab_sciml	16.22% <0.00%> (-0.05%)	⬇️
petab_sciml_benchmarks	14.80% <0.00%> (-0.04%)	⬇️
python	70.44% <95.08%> (+0.04%)	⬆️
sbmlsuite-jax	?

Flags with carried forward coverage won't be shown. Click here to find out more.

Files with missing lines	Coverage Δ
python/sdist/amici/adapters/fiddy.py	91.47% <95.08%> (+1.91%)	⬆️

... and 6 files with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.