[FonaTech]

0.3.3a1

This PR closes the gaps left by 0.3.3a0. Kernel defaults are unchanged.

Planner

A looped plan (recurrent_depth > 1) now also recommends six more loop knobs. Each comes with a sources entry that gives the 15M measurement behind it:

• recurrent_later_topk=0
• recurrent_later_groups=0 (never 1)
• lambda_iteration_cluster=0
• lookahead_teacher_scope="first"
• recurrent_dt_min=0.05
• recurrent_dt_max=1.0

The Web UI shows these as planner recommendations, and ui/presets_v3.json is regenerated.

Auto-tune

• All nine loop knobs are searchable.
  • recurrent_later_groups is offered as 0 or 2, never 1, and only in block order.
  • The loop probe now has 4 block layers.
  • A drawn dt_init is clamped into the drawn [dt_min, dt_max].
  • Each trial logs its median step time.
• Fix: the probe score included the weighted regularizers. The probe now trains on chronos_loss_term and scores cross-entropy.
• Fix: lambda_lookahead never reached the probe. With 1e-3 and 1.0, scores now differ after the first update.

Web UI

• Fix: Field.as_dict dropped 0 and 0.0 because 0 == False, so 48 fields lost min=0. It now drops only unset values.
• Help text gives the measured reasons in en, zh-Hans, zh-Hant and ja.

Translation demo

• models/chronos-enzh-50m-sft and its driver test now live under demos/translation/. CI runs the driver test as a separate step.
• The main READMEs name the demo in one pointer paragraph. The measurement conditions moved to the demo READMEs.
• test_task_models_and_tests_live_under_demos checks both.

Verified locally

• Full suite on macOS (Apple silicon, Python 3.11, MLX on CPU): 728 passed, 4 skipped. ruff check . passes.
• The demo driver test (15 items) passes from its new folder.
• Wheel and sdist both carry the general_32k tokenizer. Neither carries demos/, a model folder or weights.
• End to end through the Web UI auto-tune endpoint (in-process client, no port): 4 trials × 10 steps searched all nine loop knobs plus lambda_lookahead. Every trial completed, and the best patch carries every searched knob.

Open

• The loop-knob recommendations rest on one seed at 15M.
• Lazy block-order serving is not measured.
• Part of the extra step time of looped training is still unattributed.

🤖 Generated with Claude Code