Seed the Fortran reference so parity runs are reproducible

[neuromechanist]

Makes the Fortran reference reproducible, which #228 assumed impossible. Stacked on #233.

We already had the fix, just not in the tracked source

sccn/amica PR #54 — portable + seedable random_seed — has been applied at build time by native/patch_sources.py since epic #165, with the tracked amica15.f90 deliberately left as a read-only mirror of upstream master. Upstream has not merged #54, so the reference we validate against has been the unseeded one.

Per maintainer direction, pamica now carries #54 in its own source rather than waiting: an unseedable reference cannot anchor a parity gate, so mirroring upstream costs more than it buys. The patch script keeps working on an unpatched upstream source and is now normally a no-op.

Measured

Bundled sample, 50 iterations, two runs each:

configuration	max abs diff in W
legacy amica15mac, unseeded	0.42 – 0.59
native build + seed 42	3.3e-05
native build + seed 42 + max_threads 1	0.0

Seeding pins the initialization; the residual is thread reduction order. Both must be pinned, so run_fortran_amica now sets seed (from its own seed argument, which it previously accepted and ignored) and max_threads 1 as overrides rather than reading them from params.

End-to-end through the harness: two reference runs are now bit-identical.

A trap this opened

native/validate_shim.sh calls patch_sources.py --pin-seed. With the source already patched, patch() sees its marker and returns early, so --pin-seed would have silently done nothing and the shim-vs-mpif90 comparison would have run non-deterministic builds. --pin-seed now edits the file after patching, and fails loudly if its anchor is missing.

A bug my own unit tests missed

The first cut wrote field_dim [30504] — Python list repr — and the Fortran parser aborted at read time. Every unit test passed because they all used hand-picked keys; files and field_dim are lists in the real sample_params.json. Caught only by running the harness.

There is now a test that feeds the actual sample_params.json through and asserts no value keeps list syntax, plus one for space-separated list rendering.

Compatibility

The legacy amica15mac remains the default binary and still runs: the Fortran parser has no case default, so it silently ignores the seed keyword. Verified.

What this unblocks

#228's remaining question was which of four gate designs to adopt given a non-reproducible reference. With the native engine the reference is now bit-reproducible, so a genuine parity gate is available — the choice becomes whether to make the native engine the harness default. That decision is not in this PR.