[anwesh2410]

What

Adds __post_init__ validation to FitConfig and EmbeddingConfig so invalid
values are rejected at construction time with a clear message.

Why

Today these accept out-of-range values silently:

• FitConfig(target_teacher_agreement=1.5) / -0.2 / 0.0
• empty frontier_targets, or values outside (0, 1]
• min_deploy_coverage outside [0, 1], max_fit_labels=0, retrain_every=0
• EmbeddingConfig(batch_size=0)

A bad value either silently skews the fit or surfaces much later as a confusing
error far from its cause. Since teacher-parity is the core guarantee, the knob
that controls it should fail fast and loud.

Changes

• FitConfig.__post_init__ / EmbeddingConfig.__post_init__ raising ValueError
  with specific messages, e.g. target_teacher_agreement must be in (0, 1], got 1.5.
• 5 tests in tests/test_fit.py covering rejection + that valid configs still build.

Compatibility

Non-breaking — every valid config (including all existing tests and the demo)
constructs exactly as before. Full suite green locally (26 passed).

[adrida]

Validated and merging. Every bound matches the field's real default and semantics: target_teacher_agreement in (0, 1], frontier_targets non-empty with each value in (0, 1], min_deploy_coverage in [0, 1], max_fit_labels > 0, explore_rate in [0, 1], retrain_every > 0, min_labels_to_start >= 0, batch_size > 0. Neither dataclass had a pre-existing __post_init__, so there's no interaction with FitConfig.embedding's default_factory. Additive, CI green across 3.9 to 3.13, and it merges cleanly on top of the two PRs that just landed.

Reasoning for the fail-fast is right: teacher-parity is the core guarantee, so the knob that controls it should reject bad input at construction with a clear message rather than skewing a fit silently or erroring far from the cause. Thank you.

One note for the record, tied to your #6: explore_rate, retrain_every, and min_labels_to_start are the three fields that issue asks about. I'm keeping them (they parameterize the online loop, more in my reply there), so validating them here is exactly what I want.