[clkao]

calibrations: add the zh→en translation-head calibration for the Hunyuan 8bit model

Summary

Upstream ships one translation-head calibration for
mlx-community/Hy-MT2-1.8B-8bit: en→zh. The simultaneous-MT engine loads a
calibration at construction and raises when the file for the requested
direction is missing, so --simultaneous currently works for en→zh only.

This PR adds the zh→en calibration file in the same format, produced with the
same procedure as the en→zh file: alignment-head detection over translation
datasets, translation-score (TS) ranking, stability checks, and a promotion
gate. The heads were detected with the PyTorch bf16 reference
implementation and transferred to the 8bit MLX checkpoint; the transfer was
verified (the 8 calibrated heads are the top-8 by TS score on the MLX
checkpoint, and the attention matrices agree with the PyTorch eager
baseline). The file records the full provenance: pinned model revision,
quantization, source hashes, per-check stability flags, and the promotion
gate result.

User impact

• --simultaneous with zh→en on Hy-MT2-1.8B-8bit works on a fresh
  install. Without this file the engine raises at startup.
• No existing behavior changes. The en→zh file is untouched; the loader,
  the engine, and the commit policy are unchanged.

A question for the maintainer (separate decision)

On main, a missing calibration raises at engine construction. An earlier
design used silent deactivation instead: log a warning naming the missing
(model, direction) tuple, disable the simultaneous path, and fall back to
the base translate-on-close behavior. The simultaneous AlignAtt paper
hard-fails in the same situation. Both are defensible; the question is which
the project wants. This PR ships files only and does not change the behavior
either way.

Calibration quality notes

• Top head: layer 9, head 5 (TS score 0.77 on the 8bit MLX checkpoint,
  0.79 on the PyTorch bf16 reference), the same primary head as the en→zh
  and ja→zh calibrations, which supports these being general alignment
  heads rather than direction-specific ones.
• 8 heads total, matching the en→zh file's head count.
• The test suite includes a golden-attention test: the heads are loaded
  from the bundled file and must reproduce the recorded commit decisions of
  the calibrated engine on real captured attention slices. A corrupted or
  reordered head list fails the test (verified with top-1 and top-3
  corruption).

Validation

• New tests (tests/test_simul_calibration_files.py): the file loads
  through load_calibration with an exact prompt match; the engine
  constructs for zh→en where it raised before (negative control: a missing
  file still raises); the golden-attention reproduction test; the
  corruption-detection test.
• uv run pytest -q tests/ --ignore=tests/test_pipeline.py: same result as
  main apart from the new tests.
• ruff check .: passes.
• Live smoke on Apple Silicon (Qwen3 ASR + Hy-MT2-1.8B-8bit, zh→en,
  simultaneous): drafts release during speech; draft coverage 0.83 over four
  finals with the paper commit policy.

Checklist

• I searched for an existing issue or discussion and linked it when
  relevant. (Will link the display/event-stream issue if the maintainer
  prefers calibrations discussed there.)
• I added or updated tests for behavior changes.
• I updated documentation for user-facing changes. (The file itself
  documents provenance; no user-facing docs change.)
• I ran ruff check ..
• I ran the relevant pytest suite.
• I did not commit credentials, model weights, generated caches, or
  private data.