Commit PLM parity fixtures and gate parity tests on FERRITIN_HF_TESTS (ferritin-100.2)

[zachcp]

Closes ferritin-100.2 (P0 under the ferritin-100 PLM-hardening epic).

What this does

Establishes the correctness floor for the PLM parity suite: commits the small ESM2 + AMPLIFY golden-fixture references and makes the parity tests actually run.

• Commit fixtures — esm2_parity.safetensors + amplify_parity.safetensors (8.5 KB combined, budget was ~2 MB), un-gitignored. Generated from canonical HuggingFace models.
• Un-#[ignore] the parity tests — gated instead on FERRITIN_HF_TESTS=1 via support::parity::hf_tests_enabled. A plain cargo test exercises the wiring and skips the multi-hundred-MB weight download; the env var enables the real numeric comparison.
• Nightly workflow — .github/workflows/nightly-parity.yaml (cron + workflow_dispatch, HF download cache) sets FERRITIN_HF_TESTS=1 and runs cargo test -p ferritin-plms -- --include-ignored.
• Fix generate_amplify_fixtures.py for modern transformers (trust_remote_code, AutoModel per the repo's auto_map, additive attention mask, materialize the meta-device RoPE buffer).

Fixtures immediately caught real bugs (filed separately)

Validating the committed fixtures against the Rust ports surfaced exactly the silent numerical drift this epic exists to expose:

• ferritin-n5g (P1 bug) — ESM2 Candle logits diverge from HF on identical bare-token input (argmax differs; AA logits off by 9-15). Confirmed, no confounder.
• ferritin-edm (P2 bug) — AMPLIFY 120M distribution divergence (argmax matches; softmax off by ~0.49 on one sequence). Needs confirmation with real xformers / on GPU.
• ferritin-100.7 — run_forward adds no BOS/EOS, so the test's L+2 alignment is wrong; folded into the special-token-contract issue that owns it.

The nightly is expected to run red under FERRITIN_HF_TESTS=1 until those are resolved — that is the intended "make drift visible" signal.

Test plan

• cargo test -p ferritin-plms — green (111 + integration tests); parity tests run and skip cleanly without the env var.
• Nightly / local FERRITIN_HF_TESTS=1 cargo test -p ferritin-plms — runs the real comparison (currently fails per n5g/edm, as designed).

Notes / deviations

• Fixtures live in crates/ferritin-plms/tests/fixtures/ to match the ParityFixture::load harness finalized in ferritin-100.1, rather than the issue's original ferritin-test-data suggestion.
• ProteinMPNN parity fixture not committed / test left #[ignore]d: its generator needs the external dauparas/ProteinMPNN repo on PYTHONPATH, out of scope here.

🤖 Generated with Claude Code