refactor(desktop): push the transcript diagram theme through a Cmd

[tonyfettes]

Why

#1379 rendered the effective theme as data-transcript-theme on the transcript root and watched that attribute with the content MutationObserver so Mermaid diagrams could re-render on a theme switch. That is the DOM used as a message bus between the root and the component's subscription, the pattern the last few transcript PRs have been removing.

Rabbita never re-runs a component's subscriptions on an input-only change ("Changing the input alone does not send a message, run update, or refresh subscriptions"), so the theme cannot reach the diagram mount through Input or the subscription payload either.

What

• Theme::apply batches @transcript_component.apply_theme(dark) next to the existing @fileeditor.apply_theme, the same Cmd-push shape the file editor already uses for an imperative widget outside Model.
• The component package keeps a package-level transcript_theme: the current value plus a registry of live transcript mounts. A mount registers in its constructor and reads the current value there, unregisters on dispose, so the command and the first mount may run in either order. Only Mermaid is re-rendered; Diago SVGs follow the palette through CSS variables.
• Removed: Input.dark_mode and its three call sites, the rendered data-transcript-theme attribute, the attribute_filter entry, the get_attribute read-back, and the per-diagram mut dark_mode.
• The e2e now asserts the root data-theme projection written by the theme command instead of a component-rendered attribute.

Verification

• moon check --target js
• moon test --target js for frontend/transcript/component (21) and frontend (467)
• transcript_diagrams.spec.js in Chromium: 6 passed, including the dark-mode re-render case
• pkg.generated.mbti synced from the js _build artifact; the diff is exactly the removed dark_mode? and the added apply_theme

🤖 Generated with Claude Code

https://claude.ai/code/session_018f26rE1L4kzBf5ddeqn99i