feat(imf): Interscript Model Format v1 — spec, validator, packer

[ronaldtse]

Summary

Work order TODO.runtime-arch/01 (rababa): the Interscript Model Format v1 —
the versioned, portable artifact of interscript-ml. A model.zip adoptable on
its own terms, like ONNX itself: zip + sha256 + two ONNX sessions, no training
code required.

Three guarantees baked into the format:

1. One runtime everywhere — tokenizer is raw UTF-8 bytes only (pad=0,
   EOS=1). No vocab files, no sentencepiece. Non-byte models enter via
   distillation (WO07), never a second tokenizer system.
2. Old runtimes load it — opset pinned to 14 (Ruby onnxruntime gem's
   bundled ORT is old; verified the hard way in secryst PR #44).
3. Every number traceable — metrics cite a RESULTS.md anchor; the strict
   gate refuses zips without parity + anchored metrics.

What's here

• docs/imf-v1.md — full schema, validation levels, CLI usage
• src/imf/ — metadata schema, validator (base load-time checks + strict
  release gate), packer (computes sha256, self-validates), CLI:
  python -m imf validate|info|pack
• tests/test_imf.py — 20 specs over real zips with real ONNX graphs
  (tiny graphs built via the onnx package; no doubles)
• CI: pytest job installs export extras so graph contract checks run
• models/khm-latn/khm-latn-1.0-fp16.zip upgraded from the secryst PR #44
  zip via imf pack — base-valid (strict correctly pending the WO03 parity
  gate). Pack sources (metadata.yaml + README.md) committed; the zip
  itself is gitignored (769 MB).
• License normalized to BSD-3-Clause across the repo (code and weights);
  LICENSE file added (was MIT placeholders).

Found along the way

• The fp32 khm zip on secryst-checkpoints:/khmer_byt5/khm-latn-byt5.zip
  has a CRC error in encoder.onnx on the volume itself — exactly the
  corrupt-download failure class the sha256-on-load rule exists for. It must
  be re-exported by the WO02 pipeline from /khmer_byt5/run-001/best; the
  fp16 zip is intact and validates.

Test plan

• pytest tests/ — 57 passed (20 new IMF specs)
• ruff check src tests clean
• python -m imf validate on the repacked fp16 zip: OK (base)
• python -m imf validate --strict correctly fails pending parity
• python -m imf validate on the corrupt fp32 zip: caught (CRC mismatch)
• mypy runs in CI (strict mode configured repo-wide)