fix(finetune): the Apple Silicon and research scripts fit temperatures outside the runtime clamp

[Bruce-Yii]

What

#642 aligned the Kaggle fine-tuning notebook with the runtime temperature bounds, but two other documented training entry points still clamp fitted temperatures to 0.1..10.0:

• notebooks/laya_finetune_typed_decisions_mps.py
• research/scripts/finetune_single_device.py

This PR switches both to TEMP_MIN / TEMP_MAX from laya.common.

It also extends the existing calibration regression so the supported fine-tuning fitters are checked against the same runtime bounds instead of testing only the Kaggle notebook.

Why

laya.Agent clamps stored temperatures again when loading a checkpoint.

On current main, the two remaining scripts can fit values outside the runtime range:

fitter	peaked	dipped	served
Kaggle notebook	5.0	0.5	unchanged
Apple Silicon script	10.0	0.2848	5.0 / 0.5
single-device research script	10.0	0.2848	5.0 / 0.5

That means the calibration measured during training can differ from the temperature actually used at inference time.

Using the shared constants keeps fit and serve on the same contract.

How verified

Rebased onto current main (fa9a2a7). Run from the repository root, in a virtualenv, on CPU with no downloads:

python tests/test_calibration_persistence.py      # Ran 7 tests ... OK
python tests/test_packaging.py                    # 131 passed, 0 failed
ruff check laya/ --select=E9,F63,F7,F82,F401,F811 --line-length=120   # All checks passed!
python -m compileall -q laya/ tests/              # no output

The regression exercises both ends of the allowed range. Reverting the two production changes reproduces the mismatch: the Apple Silicon and research fitters reach the wider training bounds while the already-fixed Kaggle fitter remains within the runtime range.

The rebased test file also preserves the records_from_labeled loaded-agent regression added in #832.

GitHub Actions: 23/23 visible checks passed on head 45c94ec.

Scope

No training behavior changes beyond aligning the fitted-temperature bounds with the existing runtime contract. No API or documentation changes are required.

[NandhaKishorM]

Merged.

Two PRs fixed this and I took yours for the test, not the fix. The one-line change is identical in both: TEMP_MIN / TEMP_MAX from laya.common instead of the legacy 0.1..10.0 in the MPS notebook and the single-device script. @Swatantra-66's #885 adds two hardcoded tests, one per script, each exec-ing the extracted fitter.

Yours walks notebooks/ and research/scripts/ and holds every file whose fit clamps exp() of a log-temperature, with a count assertion so a fitter that stops clamping is reported rather than passing unnoticed. On the merged tree that finds all three:

notebooks/laya_finetune_typed_decisions_2xT4_kaggle.ipynb
notebooks/laya_finetune_typed_decisions_mps.py
research/scripts/finetune_single_device.py

That matters because the bug being fixed twice is itself the symptom. #887 opens with it: "Fixes have to be made once per copy: #851 and #885 were the same temperature-clamp bug in two of them." An enumerating test would need a third entry when a fourth script appears; yours would catch it.

I mutation-tested the gate rather than trusting it. Reverting one fitter to 0.1, 10.0:

AssertionError: 10.0 not less than or equal to 5.0 :
  notebooks/laya_finetune_typed_decisions_mps.py: fit_temperature(peaked) fitted 10
FAILED (failures=1)

Red, and the message names the file and the fitted value, so whoever breaks it does not have to go looking. Restored, it is OK at 7 tests.

Reading the notebook's %%writefile cells rather than the whole JSON is the detail that makes this work on a notebook at all, since the executed source is only in those cells.

This completes the #642 parity fix across all three documented entry points. #899 landed in the same pass and deliberately left these scripts alone to avoid conflicting with you, so both are in.

Thank you.

[sang0907791651-coder]

Hhh