[kvnloo]

Implements #260.

Changed files

• tests/test_train_next_action.py (new): hermetic pytest coverage using only NumPy, pytest, stdlib, and z0int; episode/report files remain under tmp_path.
• src/z0int/train.py: two minimal input-boundary fixes, with no refactor:
  • test_load_xy_ignores_blank_lines: skip blank/whitespace-only JSONL lines before json.loads to avoid a JSONDecodeError.
  • test_main_too_few_episodes: reject 1–100 valid episodes with SystemExit("need more than 100 episodes") before the 100-row confirmation floor would leave an empty training split.

Acceptance checklist mapped to tests

• test_hash_is_unit_norm_and_deterministic — 96-dimensional norm/determinism/case-folding and 8-dimensional override.
• test_hash_short_and_empty_text — empty, None, and two-character text each produce one normalized trigram bucket.
• test_ridge_fit_recovers_identity — unregularized identity fit and output shape.
• test_load_xy_skips_unknown_family_and_joins_prev — unknown/missing family skip; family IDs, missing user/prev, and joined previous features.
• test_load_xy_no_valid_rows_exits — both empty file and unknown-only file produce SystemExit("no episodes").
• test_load_xy_ignores_blank_lines — regression for blank JSONL lines.
• test_mb_provider_unset — opt-in env unset returns a clear unavailable reason.
• test_mb_provider_unimportable_and_incomplete — missing module, incomplete provider, and fully populated fake provider resolution.
• test_main_writes_report_with_time_split — 120 alternating rows, 20/100 split, exact class counts, ridge parameters, baseline accuracy, report file/stdout equality, no MB keys.
• test_main_mb_unavailable_and_available — missing opt-in provider report and fake fit_fly/predict_fly report with n_kc=12 and seed=1.
• test_main_too_few_episodes — controlled failure with 1, 50, and 100 rows, no report written.
• No changes outside the two issue-allowed paths; no deps, lockfiles, or CI edits.
• Runtime verification of test cases (not executed in this ChatGPT connector-only work).

tests NOT RUN - ChatGPT cannot execute code
tests NOT RUN (written via connector)

Exact test command

pip install -e '.[test]' && python -m pytest -q tests/test_train_next_action.py

No CI/test results or device evidence are asserted.