Settings menu: minor UX and state-consistency bugs

[QueryPlanner]

What

Found during a code-reduction audit of the Telegram /model//thinking settings UI (now in telegram/settings_menu.py after #161). Four small, non-crashing bugs.

1. Misleading "Updating settings…" toast on pure navigation (settings_menu.py:240): answer_callback_query(query.id, text="Updating settings…") fires unconditionally before dispatch, including for action == "thinking" (just opening a submenu) and action == "back" — both write nothing, so the toast is inaccurate.
2. Redundant profile reload after a write. update_inference_profile(...) (lines 246, 273, 283) already returns the freshly merged, validated InferenceProfile, but the return value is discarded and _edit_model_menu immediately re-reads via _load_profile, causing an avoidable extra storage round-trip on every settings change.
3. Current-selection checkmark matched by rendered label text, not the underlying value (settings_menu.py:184): text=f"{label}{' ✓' if label == current else ''}". If _REASONING_LABELS is ever edited so two different effort values render the same label, the wrong option (or none/multiple) gets marked as selected.
4. Latent UI mismatch when a model's reasoning becomes mandatory (settings_menu.py:434): when reasoning.mandatory is true, "none"/"Off" is silently excluded from the option list, but if the currently stored effort for that chat is "none", the "Current: Off" line still renders even though there's no matching selectable button — a confusing but non-crashing state mismatch.

Priority

Low — cosmetic/UX papercuts, no data loss or crash risk.

Level of Effort

Small (S) — each is a localized, few-line fix independent of the others.

Sources

• telegram/settings_menu.py
• tests/test_telegram_settings_menu.py (existing coverage to extend)

Passing criteria / definition of done

• handle_callback sends a toast only for actions that actually mutate stored preferences (model/reasoning/reset), verified by a test asserting no "Updating settings…" text for action in {"thinking", "back"}.
• _edit_model_menu/callback handling renders directly from the profile returned by update_inference_profile instead of re-fetching; a test asserts _load_profile (or the storage mock) is not called an extra time after a successful write.
• The reasoning-menu checkmark match is keyed on the option's underlying value, not its label; a test constructs two options that would render the same label and asserts only the correct one is checked.
• When the currently-stored effort is excluded by mandatory=True, the menu either falls back to a valid current value or surfaces an explicit note instead of silently showing a value with no matching button; a test covers this case.
• pytest tests/test_telegram_settings_menu.py -q passes with the new assertions.
• ruff check, ruff format --check, and mypy src/blacki/telegram/ all pass with no new warnings.