docs: fix five docstring/code mismatches

[kgdunn]

Summary

Docstring-only cleanup: five spots where the prose had drifted from what the code actually does. No behaviour changes.

Fixes

1. src/scorepilot/schemas.py:300-307 — FitModelRequest: the old wording claimed n_components was used as an upper bound when auto_components is set. It isn't — the auto-fit ceiling is the fixed internal _AUTO_MAX_COMPONENTS = 20 in api/models.py, and the request's n_components is ignored. Rewrote the sentence to say so.

2. src/scorepilot/core/profiling.py:142-158 — suggest_transform: docstring implied "milder skew → signed power". The code actually returns POWER only when skew > 1.0 and the log preconditions (positive + wide range) fail; mild skew (≤ 1.0) returns NONE. Reworded to "a strong skew that isn't positive with a wide dynamic range suggests a signed power (root)".

3. src/scorepilot/core/transforms.py:51 — apply_transform Notes: claimed power defaults to c1 = 0.5, but the signature default is c1 = 0.0, which _transform_values remaps to 0.5 internally. Clarified the note to state the parameter default is 0.0 and is treated as 0.5 (signed root) inside _transform_values.

4. src/scorepilot/db/models.py:47-54 — Model class docstring listed the persisted fitted arrays as "P, W, means, scales, …". _pack_params in api/models.py (lines 51-59) only stores x_loadings, explained_variance, r2_cumulative. Updated the list to match.

5. src/scorepilot/api/models.py — replaced stale "no longer in memory" / "no longer loaded" wording at four sites (get_model ~L336-337, update_model ~L386, model_contributions ~L427, model_cross_validation ~L489) with "no longer available" / "the source dataset record is missing/deleted". Datasets are persisted now, not in-memory.

PATCH version bump (0.22.0 → 0.22.1); lockfile intentionally not refreshed (harness rule).

Test plan

• uv run ruff check .
• uv run pyright
• uv run pytest
• Skim each edited docstring in a rendered view to confirm phrasing reads naturally.

---

Generated by Claude Code