Refactor toolkit registry

[windchargerj]

Toolkit registrations and discovery results now have separate owners.

What Changed

• Added ToolkitScanner for host discovery, executable lookup, version probing, latest successful scan snapshots, and per-host failures.
• Added ToolkitRegistry for registered toolkits and the default toolkit. It refreshes matching entries by physical location, updates availability, restores backends, and serializes persisted state.
• ToolkitManager now coalesces concurrent scans, cancels a project's scan when the project closes, and exposes project-aware visibleToolkits, registeredToolkits, and registeredToolkit queries.
• Successful scans update metadata and availability but do not delete user registrations. Failed host scans retain the previous snapshot; failed version probes retain the executable but mark it unavailable.
• Persisted state normalizes legacy local and WSL/SSH host IDs to backendId, removes duplicate registrations at the same physical location, and preserves the default toolkit where possible.
• Settings, the project wizard, run configurations, build execution, and metadata consumers use the project-aware queries.
• Toolkit selectors use a live model that preserves selection, displays availability, and safely refreshes an open JetBrains popup.
• Removed the legacy global mutation APIs and ToolkitChangedNotifier.

The resulting state flow looks like this:

flowchart TB
    Hosts["Platform hosts"] --> Scanner["ToolkitScanner"]
    User["User registrations<br/>and default toolkit"] --> Registry["ToolkitRegistry"]

    Scanner -->|"latest scan snapshot"| Manager["ToolkitManager"]
    Registry -->|"registered state"| Manager
    Manager -->|"refresh by location<br/>and availability"| Registry

    Manager -->|"project-aware queries"| Consumers["Settings, wizard,<br/>run configurations, build,<br/>metadata actions"]
    Manager -->|"ToolkitListener"| Selectors["Live toolkit selectors"]

Loading

The scanner observes installed hosts and executables, the registry owns the user's registrations and default choice, and the manager combines both into the project-aware state consumed by actions and selectors.

Why

• A registration represents a user choice; a scan represents an observation that can be incomplete or fail. Discovery should report the environment, not silently rewrite user configuration.
• A complete executable scan can mark missing installations unavailable, while a transient host failure can retain the last known snapshot. This distinguishes a real environment change from a temporary probe failure.
• SSH host visibility depends on project configuration and credentials, so toolkit availability cannot be represented by one global list.
• Separating scanner, registry, and orchestration makes the state rules explicit: the scanner reports what exists, the registry owns user choices, and the manager coordinates updates and notifications.

Impact

Discovery updates availability without overwriting registrations. Consumers receive live, project-aware toolkit state while retaining the user's selection.