Three changes against the user feedback "der text ist zu
zerhackt, die filter-leiste nimmt zu viel sicht weg, llm-suche
wenn das geht":
1. Today-Hero copy rewritten in flowing prose. The fallback
story drops the code-fence templates ("`mcp.<server>.<tool>`")
and reads as a single arc per language. Same trust model,
gentler register.
2. Filter row collapses behind a single `_FilterButton` with a
"Filter · N active" count badge. Status / Source / Installed
chips now live inside `_FilterDialog` — invisible until the
operator wants them, recovers a full row of viewport. The
category strip stays inline because category is the most
common cut and benefits from being one click away.
3. New `_AskBar` replaces the old single-line `_SearchField`:
- Multi-line input (1-4 lines auto-grow) so questions don't
overflow horizontally.
- Cmd+Enter / Ctrl+Enter submits; bare Enter inserts a
newline so the operator can write multi-line questions
naturally.
- Heuristic question detector (ends in `?` or ≥4 words)
routes the submit to the System-AI's askAi RPC instead of
the substring search. Live keystrokes still keyword-search
for short queries; question-shaped input holds the grid
steady until submit so the visible result set doesn't
wipe with every space.
- The LLM is asked for a strict-JSON ranking of up to 5
modules with a one-phrase reason each. Unknown module
names are dropped silently — operators must not see
hallucinated entries.
- The new `_AiAnswerCard` renders the LLM's answer plus the
ranked match list above the grid; the grid filters to
just those names so the answer and the visible cards
stay coherent. Clearing the question collapses everything
back.
- Falls back to plain keyword search when System AI is off
(with an explanatory hint instead of the AI hint).
The AI path uses the System-AI the operator already configured
in Settings; same privacy mode, same audit-log trail. Nothing
new leaves the hub.
Signed-off-by: flemming-it <stefan.a.flemming@googlemail.com>
c7869870c4