feat(imf): WO02 ONNX export pipeline — KV decoder, fp16/int8, Modal runner

[ronaldtse]

Summary

Work order TODO.runtime-arch/02: the ONNX export pipeline that turns any
HF byte-level seq2seq checkpoint into IMF v1 zips — fp32 + fp16 + int8,
with a KV-cache decoder as the default artifact and the plain decoder as
fallback. Generalizes secryst PR #44's scripts/export_onnx_byt5.py.

Stacked on #2 (IMF v1 spec + validator). Base is feat/imf-v1; retarget
to main when #2 merges.

Hard-won export findings (all encoded in code + comments)

• Tracing mutates the traced module. torch.onnx.export leaves the
  model's runtime behavior subtly altered; the WO03 parity harness needs
  the reference pristine, so export_graphs traces a deepcopy.
• KV cache design. decoder-kv.onnx caches self-attention only.
  Cross-attention K/V are a deterministic projection of
  encoder_hidden_states — caching them breaks the attention-mask length
  invariant for any fed past length, and EncoderDecoderCache.is_updated
  is a Python bool baked at trace time. Recomputing cross K/V per step is
  correct for all past lengths and removes all cross-cache bookkeeping
  from the three runtimes. Step 0 feeds zero-length self pasts.
• fp16 must keep LayerNorm/softmax math in fp32. ORT's session-time
  SimplifiedLayerNormFusion crashes on half-converted LN subgraphs
  (InsertPrecisionFreeCast name mismatch). Blocking the LN decomposition
  ops keeps graphs loadable with default session options (old Ruby ORT
  included). Weights — the bulk — still halve.
• int8 must be MatMul-only dynamic quantization. Quantizing more ops
  inserts casts that hit the same fusion bug; quant_pre_process pins
  concrete example shapes into DynamicQuantizeLinear buffers.

Verification

• Fixture (tiny random T5, seed 42) exported end-to-end locally:
  • ONNX plain == ONNX KV == untraced torch reference, token-exact,
    4 probe strings
  • fp16 and int8 zips decode token-identical to fp32
  • all three zips pass imf validate (kv contract, opset 14, sha256)
• tests/test_imf_export.py (4 specs) — full suite 61 passed, ruff clean
• New CI job export-fixture installs train+export extras and runs the
  torch/ORT parity specs

Modal runner

src/gpu/modal_export.py — CPU-only (exports never compete with A100
training), watchdog invocation documented:

until modal run --detach src/gpu/modal_export.py --model khm-latn; do sleep 60; done

Registry: khm-latn, urd-g2p, urd-diac → secryst-models:/imf/<model>/.
The khm-latn run re-exports fp32 from scratch, replacing the CRC-corrupt
zip found on the checkpoint volume.

Test plan

• pytest tests/ — 61 passed locally
• ruff clean
• CI green (incl. new export-fixture job)
• Modal exports for khm-latn / urd-g2p / urd-diac complete + validated
  (running next; results in this PR)

[ronaldtse]

Root cause found for the decode convention: the stock byt5 tokenizer maps byte b -> id b+3 with a trailing EOS — the helpers (and PR #44's Ruby engine) fed raw bytes. Fixed in 83c163d (canonical table in the spec + helpers + tests); khm golden now decodes exactly ('រោក' -> 'rok', 'សង់ផ្ទះ' -> 'sangphteah'). Graphs were always correct — no re-export; parity gates re-running with the fix.