Gray out Train Aligners until a dataset has manual alignments

[BeckettFrey]

Split out of #163 so it can be reviewed on its own.

What

Disables the Train Aligners pipeline step — both the sidebar entry and the page behind it — whenever no registered dataset has a manual/hand alignment. It re-evaluates on reload(), so registering a qualifying dataset enables the step without restarting the app.

Why

Training only produces a meaningful model when it can learn from a hand-corrected alignment. Training against machine-generated alignments alone just reinforces whatever errors those alignments already contain.

Concerns with this approach

1. The page can't be opened at all, even to look at it. Disabling the stacker as well as the sidebar entry means a user cannot explore what Train Aligners is before they have the data to run it. Exploration is exactly what a first-time user wants from the pipeline sidebar.

2. It gives this one step behavior no other step has. Every other sidebar entry is always reachable. Train Aligners becomes the sole step whose availability depends on external state, which is a special case in PipelineFormStack (_menu_items, _refresh_training_availability) to maintain and reason about from here on.

3. It never says inline how to enable it. A grayed-out row with no reachable page gives the user no in-place path forward. The explanation lives in the Pipeline Overview under stage Ⓑ and on the Train Aligners page — but the page is precisely what they cannot open, so one of the two places explaining the requirement sits behind the gate it explains.

Counter-proposal: filter the alignment dropdown instead

Leave Train Aligners always reachable, and make TrainingStacker's alignment dropdown list only hand/corrected alignments, excluding machine-generated ones.

The machinery already exists. on_dataset_changed (training_stacker.py:84) populates the dropdown from alignments.list_alignments(...) and already calls alignments.get_alignment_type(a) for display — which returns "hand", "corrected", or "automatic". The filter is that predicate applied one line earlier. The empty state exists too: when a dataset has no usable alignments the dropdown disables itself behind a placeholder (training_stacker.py:109-114), so "No hand-corrected alignments for this dataset" lands where the user is already looking, at the moment they pick the dataset.

Same guarantee — you cannot train on generated-only data — while:

• keeping the page browsable for exploration,
• keeping every sidebar step uniformly reachable, with no per-step availability logic,
• putting the message inline at the point of failure instead of a page away.

It is also the smaller change: no QListWidgetItem tracking, no item:disabled styling, no any_dataset_has_manual_alignments(). And it removes a second source of truth — the gate keys on the dataset's hand_alignments_path, while the dropdown keys on each alignment's alignment_type.

Changes (as currently implemented)

• storage/datasets.py — any_dataset_has_manual_alignments() helper.
• gui/pages/pipeline/__init__.py — track menu items by step id; _refresh_training_availability() on init and on reload.
• gui/__init__.py — QListWidget::item:disabled styling so the grayed-out step reads as disabled.
• gui/pages/pipeline/training_stacker.py — explanatory note on the page.
• config/profiles/explanatory/pipeline_definitions.yaml — Pipeline Overview copy for stage Ⓑ.

Testing

Manual: with no hand-aligned dataset registered the step is grayed and unclickable; after registering one and reloading, it becomes available.

Related

• MVP polish: W2TG default, tier order, app branding #163 — the original combined PR, closed in favor of these splits
• Ship another application icon in packaged builds #169 — application icon (open); Add university credit line to the first-launch splash dialog #170, Render the word tier above the phone tier in TextGrid timelines #168, Default W2TG "Use Speaker Adaptation" to on #167 — merged

[BeckettFrey]

@nrgslp see the concerns section and counter-proposal above. This starts to touch on architectural decisions which conflict with my thinking. Happy to hop on a call, or discuss Monday. Thanks again for these changes.