muscl_eps is never broadcast: rank-divergent MUSCL reconstruction on multi-rank runs

[sbryngelson]

Symptom: any multi-rank MUSCL run with a user-set muscl_eps uses the user's value on rank 0 and the per-limiter default on every other rank — silently divergent reconstruction.

Mechanism: muscl_eps is registered and namelist-read (rank 0); the per-limiter derivation runs only under f_is_default(muscl_eps) (src/simulation/m_global_parameters.fpp:736-743 on master) and defaults are assigned on all ranks — but the user's value never leaves rank 0 because muscl_eps is missing from the real-valued broadcast list in src/simulation/m_mpi_proxy.fpp:128 (weno_eps is there).

Introduced: with the parameter itself in #1383 (02eb4ece0, 2026-05-01, @ChrisZYJ). Tagging @ChrisZYJ for context.

Fix: #1553 adds the broadcast (and the registry-driven generation in the same series prevents recurrence).

[ChrisZYJ]

I'll look into this. By the way, I originally added muscl_eps as a tunable parameter because I thought there were stability reasons for having eps and designing MUSCL the way it was. However, after looking into it a bit more, I think the textbook eps=0 is actually the standard. Having eps might be a common mistake that unintentionally makes MUSCL reduce to first-order accuracy under small amplitudes. If you agree, I would prefer to remove muscl_eps and default everything to having no eps (in a form that avoids division by zero).