Summary

Adopts the reproducible-seed version of the amica15 source (sccn/amica PR #54, supersedes #53) into pyAMICA's native binary build. The CI-built release binaries (release-binaries.yml) now support a seed option in input.param: a run's random initialization is reproducible when seed is set, clock-random otherwise.

Change

• native/patch_sources.py: the build-time seed patch now applies PR Refuse degenerate fits in the AMICA wrapper (#50) #54's version instead of Return best iterate from AMICATorchNG.fit (#51) #53's. It (1) sizes random_seed(PUT=...) via random_seed(SIZE=...) (portable across ifort/gfortran, as before) and (2) wires a seed param -> deterministic per-rank seeding when set. Four idempotent, anchor-verified patches (header decls, seed logic, BCAST, parser case).
• The tracked pamica/amica15.f90 stays the pristine parity reference (== sccn/amica master); only the build copy is patched, as before.
• validate_shim.sh's --pin-seed still works (pins the unseeded/clock branch for the bit-for-bit shim-vs-mpif90 comparison).
• README/build.sh comments updated (Return best iterate from AMICATorchNG.fit (#51) #53 -> Refuse degenerate fits in the AMICA wrapper (#50) #54).

Testing

Built the shim binary locally (Apple Silicon gfortran) and ran the bundled 32-ch sample, single-threaded (deterministic reductions):

• Two runs with seed 42: W byte-identical (LL -3.4570950991 both) -> reproducible.
• Two runs with no seed: W differs (LL -3.4571252798 vs -3.4571719932) -> clock-random.

--pin-seed applies cleanly (3 anchors, clock pinned); ruff + ty pass.

Notes

• The release binaries rebuild via release-binaries.yml on a release (or manual workflow_dispatch), not on this merge -- cut a release / dispatch to publish binaries carrying the option.
• Exposing seed through the AMICANative Python engine (pass from Python, not just a hand-written input.param) is a natural follow-up.