Demo agent: in view, legible, and quicker

[jpcamara]

Stacked on #102.

Three complaints from using it: the agent was often off screen, it was unclear how to interact with it or what it would do next, and too much time was spent waiting.

In view

• A bar pinned under the page header always shows the current status and detail, with a "jump to it" link that scrolls the editor to the agent's cursor and a follow switch (on by default) that keeps the cursor in view as it moves, held for four seconds after you type. Lexical: the cursor element is scrolled into view once drawn. CodeMirror: the awareness position is resolved to an index and EditorView.scrollIntoView does the rest.

Legible

• The agent introduces itself as its first ledger entry: what it will do first, that it works through the list under a heading that names it at the same time, and how to talk to it.
• "up next: …" before each task, with a moment to say @agent pause.
• Buttons under the bar for what you can say: ask, hand it a task, draft this section, pause, resume. Each puts the line on a new paragraph (Lexxy, through the editor's own API) or line (CodeMirror) at your caret and leaves the caret at its end.

Quicker

• StreamJob: a model stream with its own pacer, stepped by the watch loop, so several run at once and the loop keeps reacting. The review and the first task from the list start together at the invite; draft text appears two to three seconds after the invite. While both run, one combined status ("writing a review and drafting Rollout") keeps the ledger and the bar from flickering, and the caret stays with the draft.
• Quick calls, the decision after a change and answers, go to a model without a reasoning phase: AGENT_FAST_MODEL, default accounts/fireworks/models/minimax-m3 (2.4s to first token on the decision prompt, clean JSON; gpt-oss-120b also works). Reviews and drafts keep the reasoning model. Measured with the real models: a decision landed 3s after a sentence ended, against 20s before.
• A change is only weighed once it reads finished: the block ends with punctuation, or the person's caret has left it.

Verified

• Chrome, stub, both pages: draft started 3s after invite with the review still writing; ledger reads hello, reading, writing a review, up next, drafting, writing a review and drafting, wrote a review, drafted; the bar follows; jump scrolls to the agent; the Ask button makes a new paragraph with @agent at the caret and the answer arrives.
• Chrome, real models, markdown page: first text 2s after invite; review and draft interleaved; decision 3s after the sentence ended with a sensible reason.