fix(trainer): unify callback kwargs across all trainers (#802) - #1023

All review items addressed in b659674:

1. Ratchet Guard Satisfied: Retained explicit eval_gate_config=eval_gate_config across all 16 trainers and added include_eval_gate: bool = True to soup_callback_kwargs (passed with include_eval_gate=False by callers). All 14 cases in test_code_review_critical.py::test_trainer_wires_eval_gate_config pass.
2. Version Literal Single-Source-of-Truth: Dropped the version prefix from src/soup_cli/config/schema.py:6780 (#802 — ...). tests/test_issue627_unknown_config_keys.py::TestTheDeadline::test_the_version_is_written_out_in_exactly_one_source_file passes.
3. Fragment Renamed: Fragment renamed to changelog.d/0.75.0/1023.fixed.md ending with (#802 by @kok-o in #1023).
4. Cleaned batch_size fallback: Cleaned repeated fallbacks across trainers to direct self._batch_size.

All 21 tests in tests/test_issue802_callback_kwargs_in_trainers.py and full suite pass cleanly; ruff check is green.

All remaining review items addressed in c95cdfe:

1. Behavioural Wrapper-Path Tests: Added TestTrainerWrapperBehaviouralCallbackWiring in tests/test_issue802_callback_kwargs_in_trainers.py verifying that both GRPOTrainerWrapper.train() and DPOTrainerWrapper.train() genuinely forward training monitoring configuration (loss_spike_recovery, grad_accum_auto_tune, resolved batch_size) into the instantiated SoupTrainerCallback.
2. Simplified batch_size Resolution: Replaced the nested getattr(self, "_batch_size", ...) fallback across all 15 trainers with direct batch_size=self._batch_size.
3. Fragment Renamed: Renamed to changelog.d/0.75.0/1023.fixed.md ending with (#802 by @kok-o in #1023).

All 23 tests in test_issue802, 16 critical eval gate tests, the #627 version deadline test, and ruff check pass cleanly.

Merged — thank you. You closed the thing I actually asked for, in the form I asked for it, and that is what moved this.

Last round the finding was that the wiring was pinned by a source-text scan, so reverting trainer/grpo.py to a hand-built kwargs dict changed nothing any test could see. TestTrainerWrapperBehaviouralCallbackWiring fixes that properly: it drives GRPOTrainerWrapper.train() and DPOTrainerWrapper.train() with the trainer mocked, captures add_callback, and asserts the flags on the callback that actually arrives — a behavioural pin, not a signature assertion. That is not the same finding again, and I want to say so explicitly since I have given you the same note twice before.

I also liked that you kept eval_gate_config= literally in all sixteen wrappers and passed include_eval_gate=False, rather than taking the tidier route that would have broken the fifteen tests. The tidier route was available and you did not take it.

What I ran, on the tree rebuilt by merging your head into current main (GitHub's merge ref is pinned at run creation, so its green marks described a main several merges behind), base interpreter, peft 0.20.0, soup_cli.__file__ under the scratch tree:

• The 15 previously-failing tests — test_code_review_critical.py::test_trainer_wires_eval_gate_config (14 params) and test_issue627_unknown_config_keys.py::TestTheDeadline::... — 15 passed.
• Callback / auto-tuning / watchdog / consumer-guard suites — 317 passed.
• Recipes, snapshot, templates, config — 332 passed.
• Full ruff check clean, --select F821 clean. git diff origin/main...<head> -- tests/: +402 lines, zero deletions, no skip, no xfail, no widened tolerance.
• Enumerated the wrappers myself rather than trusting the description: SoupTrainerCallback( appears in exactly 16 files and all 16 call the helper once. tts and bitnet inherit from SFTTrainerWrapper, preference dispatches to the wrappers, and prm / mole_routing / unlearn attach no callback and are now refused at parse.

Mutations: reverting trainer/grpo.py and reverting trainer/dpo.py each fail exactly the new behavioural test. One mutation still survives — reverting trainer/kto.py alone leaves 51 passed, because the other fourteen are held by the source scan only. That is the first item below; it is not blocking, and it is smaller than what you have already done.

Follow-ups, none blocking, in the order I would do them:

1. Behavioural coverage is 2 of 16. Reverting trainer/kto.py to a hand-built dict passes the whole suite. The two wrappers you tested take the same object.__new__ + mock-trainer shape, so parametrizing the test over the rest should be mechanical. If you would rather not, say in the test docstring that the other fourteen are held by the source scan — so the next reader knows which is which.
2. loss_spike_recovery and grad_accum_auto_tune reach the training callback on SFT only: 15 trainers build it without them, 3 build none #802's second criterion, literally. It asks for a GRPO or DPO callback built through the wrapper, fed a loss above threshold for patience steps, to write spike_recovery.json. Your new test stops at the flags; the hint-file assertion drives the helper, not a wrapper. Driving cb.on_log inside the wrapper test would close it word for word. I am accepting Resolves #802 as it stands because the restated ask is met, but that is the gap.
3. include_eval_gate is now a dead parameter. All 16 callers pass False, so the True branch survives only in test_default_values. Drop the parameter or the default.
4. One sentence in the fragment. _validate_callback_monitoring_task_compat turns a config that used to load (task: prm + loss_watchdog: true) into a load-time refusal. I scanned recipes/, templates/, examples/ and src/soup_cli/templates/ — nothing shipped hits it — but the fragment is .fixed.md and does not mention that a key is now refused. Someone whose soup.yaml stops loading will look there first.

And one observation that is not this PR and that I will file separately unless you want it: your validator gates on task only. trainer/mlx_sft.py contains zero references to loss_watchdog, loss_spike_recovery or grad_accum_auto_tune, so on backend: mlx all three are still accepted and read by nothing — the same shape as #802 and #683, one axis over. You found the task axis; the backend axis is the other half.

Filed the backend-axis observation from my merge comment as #1069: loss_watchdog, loss_spike_recovery and grad_accum_auto_tune are read by nothing on backend: mlx, because _validate_callback_monitoring_task_compat gates on task only.

Not a defect in your PR — you fixed the task axis and that is what #802 asked for — but it is the other half of the same shape, and you found the seam. It is unassigned; say the word and it is yours, since you already have the validator's context.