Hub — run titles, filter, scrolling

Every row leads with a short title (at most 48 cells) for what the item is actually doing. For a run the engine derives it when the run is created and stores it as title in run.json: injected preambles (HARD INVARIANTS, RULES, MEMORY, PRIOR RUNS, instructions) are dropped, a labelled TASK: / RESEARCH QUESTION: / GOAL: block wins over the wrapper text, a leading "As of ," clause is dropped, and a follow-up reads follow-up <id>: <reply>. Runs recorded before titles existed get the same derivation at display time (kt4xb), and Ctrl+R overrides it per run. Two parallel implement-task runs therefore read by their distinct task, not by an identical workflow name. A chat row shows its session title.

When the list is taller than the terminal it windows to fit (the input and the hint row are never pushed off-screen), shows ↓ N more below under the last visible row and ↑ N more above at the end of the strip once rows have scrolled off the top, and edge-scrolls to keep the selection visible; PgUp/PgDn move a page. On a short terminal Done and Failed also fold sooner into … N more.

/ filters live (see "Fleet hub"): words match title, activity, folder or a run-id prefix, and the failed / needs / live / today tokens scope by status and day. ⏎ keeps the filter and returns to the list; esc clears it. The strip under the header names the active filter. The full task, facts (folder, steps, tokens, cost, id) and the last activity lines are in the peek (Space), not in a pinned card.