fix(predict): classification labels are strings whatever return_proba is (#45)

[amrit110]

Closes #45.

Problem

A classification's predictions changed type with return_proba:

	return_proba: true	return_proba: false
tabicl-v2 (before)	["1.0"]	[1.0]
tabpfn-3.5	["1.0"]	["1.0"]

Without probabilities, TabICLBackend._predict returned model.predict(X).tolist(). That keeps the label dtype, and it goes against the backend protocol, which says to derive labels from probabilities.

Fix

• New backends.base.classification_outcome(proba, classes, return_proba): argmax over [str(c) for c in classes], with probabilities and classes only when asked. TabICL and TabPFN both use it, so the two backends can't drift apart again, and TabICL no longer calls predict().
• API schema: the predict responses' predictions field documents the contract. Classification gives strings whatever return_proba is; regression gives numbers.
• Behaviour change: tabicl-v2 clients that relied on numeric labels without return_proba now get strings, as with tabpfn-3.5. This is noted in the CHANGELOG.

Tests

• New tests/unit/test_classification_outcome.py:
  • The helper's contract.
  • Both real backends' _predict, driven by a stub estimator whose predict() raises. The TabICL case fails on main's code (checked).
• Unit tests: 501 pass.

Not in this PR

• Float-form labels: "1.0" for a JSON 1 comes from train_y being parsed as floats. Whether to keep integer labels as integers is the separate question noted in TabICL classification predict returns numeric labels without return_proba, strings with it #45.
• Gateway docs: they warn about this exception (aieng-proxy-gateway#471). The warning can come out once this is deployed.

No architecture change.

🤖 Generated with Claude Code

[amrit110]

Review fixes pushed. The independent review found no bugs: tabicl 2.1.1's and tabpfn 8.5.0's predict() are exactly argmax(predict_proba) over classes_, so the labels are unchanged, and only the type changes.

• Helper: moved to backends/_outcome.py, so base.py stays protocol-only.
• Fake backend: reports through the same helper. On ties its label is now the argmax of its own probabilities.
• Real-model check: tests/integration/test_tabicl_kv_real.py now also asserts that return_proba=False gives the same labels, drawn from classes. This runs on the GPU only.
• Wording: API doc and CHANGELOG.

Unit tests: 501 pass.

[codecov-commenter]

Codecov Report

❌ Patch coverage is 95.45455% with 1 line in your changes missing coverage. Please review.
✅ Project coverage is 79.45%. Comparing base (7579da1) to head (26e28a3).

Files with missing lines	Patch %	Lines
src/tabctx/backends/fake.py	85.71%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #56      +/-   ##
==========================================
+ Coverage   79.43%   79.45%   +0.02%     
==========================================
  Files          41       42       +1     
  Lines        3209     3218       +9     
==========================================
+ Hits         2549     2557       +8     
- Misses        660      661       +1     

Flag	Coverage Δ
integration	30.08% <18.18%> (-0.06%)	⬇️
unit	75.94% <77.27%> (+<0.01%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

Files with missing lines	Coverage Δ
src/tabctx/backends/_outcome.py	100.00% <100.00%> (ø)
src/tabctx/backends/base.py	100.00% <ø> (ø)
src/tabctx/serve/app.py	35.87% <100.00%> (+0.20%)	⬆️
src/tabctx/backends/fake.py	98.14% <85.71%> (-1.86%)	⬇️

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.