ci: intersect flat --solvers list with each problem in the benchmark matrix

[andrinr]

Problem

benchmark-run.yml runs a matrix over problems and applied the same flat inputs.solvers list to every leg's mosaic run (--problems "${{ matrix.problem }}" --solvers "$SOLVERS"). A cross-domain list breaks the legs whose problem lacks some of those solvers:

Run SOLVERS="deal.II,FEniCS,Firedrake,JAX-FEM,torch-fem"
Unknown solver 'torch-fem'.
Available solvers: FEniCS, Firedrake, JAX-FEM, TopOpt.jl, deal.II
Error: Process completed with exit code 1.

torch-fem is thermal-only (structural-mesh has TopOpt.jl instead), and mosaic run (_helpers.py:471) rejects a name absent from the problem it runs against. One bad leg fails the whole dispatch — which is exactly what blocks verifying a thermal-only change (e.g. #85) whenever the matrix also includes structural-mesh.

Fix

• Add ​.github/scripts/intersect-solvers.py — prints the intersection of a flat solver list with one problem's own solver set (case-insensitive, canonical casing preserved), mirroring the existing filter-problems-by-solver.py.
• In the Run step, intersect SOLVERS with matrix.problem before calling mosaic run; skip the leg when the intersection is empty.

Each leg now runs exactly the requested solvers it actually has:

leg	requested	runs
thermal-mesh	deal.II,FEniCS,Firedrake,JAX-FEM,torch-fem	all five
structural-mesh	same	FEniCS,Firedrake,JAX-FEM,deal.II (torch-fem dropped)
(no overlap)	same	leg skipped with a ::notice::

The pull-solver-images.py step already skips out-of-problem solvers (pull-solver-images.py:96), so only the run step needed this.

Verification

• intersect-solvers.py compiles; YAML validates; the intersection logic is unit-tested (cases above).
• Full end-to-end needs a benchmark dispatch (Docker + GHCR images), which I can't run here — the matrix legs are otherwise unchanged.

🤖 Generated with Claude Code

[PasteurBot]

📊 View the full benchmark results

No benchmarks ran for this PR, so there is no status report.

---

[dionhaefner]

@andrinr Not quite understanding what bug is being fixed here. How can I trigger the error above?