Fit processors for schema fields that were not supplied

[solarsys]

Problem

SampleBuilder.fit fitted processors for a side (input or output) only when no processors were supplied for that side. Supplying a pre-fitted processor for one field therefore left every other field on that side without a processor. Its raw value, e.g. a Python list like [40.0], reached the model with no error.

That's easy to hit with the "fit on train, reuse for test" pattern: reuse the code vocabulary, forget one field, and that field is silently left raw.

create_sample_dataset(samples, {"codes": "sequence", "age": "tensor"}, {"label": "binary"},
                      input_processors={"codes": train_codes})
# master: dataset[0]["age"] == [40.0]   (a list, no processor)

Change

• SampleBuilder.fit fits a processor for every schema field that has none. Supplied processors are used as they are and never refitted.
• SampleBuilder copies the supplied processor dicts. Without the copy, filling in missing fields would also add them to the caller's dict, for example the training dataset's input_processors.

This is PR 1 of the plan in the design note on fitting processors on the training split; it's independent of the split= work.

Tests, docs, example

• New tests/core/test_partial_processors.py:

  • a missing input processor is fitted and gives tensors;
  • a missing output processor is fitted;
  • a supplied processor isn't refitted (its vocabulary doesn't grow);
  • create_sample_dataset with partial processors.

  Three of these fail on master.

• Existing processor-transfer, ignore-processor, sample-builder and schema tests pass.

• docs/api/processors.rst gets a new "Reusing Fitted Processors" section, which also documents the fit-on-train pattern.

• SampleBuilder gets a >>> example.

• New examples/reuse_train_processors.py: fit on train, reuse for test. A test-only code maps to <unk>.

• Full core suite: Ran 1401 tests … OK (skipped=76). tools/check_pr_rules.py passes.

🤖 Generated with Claude Code

[solarsys]

Closing as already merged: this PR's change (fitting processors for schema fields that were not supplied, plus tests/core/test_partial_processors.py) landed on master as part of #1273, which was stacked on it and squash-merged in e94ca1d. Nothing left to merge here.