simulate: print the job view into the terminal instead of windowing it in the alt screen

[u9g]

Opening a job in lk agent simulate windowed renderDetail inside the alt screen: instructions, transcript, and logs were paged through a viewport ~12 rows shorter than the terminal, with ↑ N more lines markers, and the terminal's own scrolling, selection, and search were unavailable for the one view that is mostly text worth reading.

The job view now leaves the alt screen and is printed with tea.Println, so the terminal owns it. The list view is untouched — it keeps the alt screen, its page windowing, and its cursor. Going back re-enters the alt screen.

How it behaves

• flushDetail runs after every message and emits only what the view gained since the last call, so growth while a job is still running (a log line, the transcript arriving with the summary) appends instead of redrawing. A render that rewrites what came before is reprinted whole, since scrollback cannot be edited in place; a resize rebaselines without reprinting, because printed text cannot re-wrap.
• Opening a job clears the screen and scrollback first, so the terminal holds one job at a time rather than every job visited this run. Erasing only the previous job is not possible: once its rows scroll off, ESC[3J is the only sequence that reaches them, and it takes the whole scrollback with it — including whatever preceded the run. That trade-off is deliberate.
• While a job is open the live region is just its status line and the hint bar, and up/down/pgup/pgdown are swallowed so they cannot move the list cursor hidden underneath.
• j and l sit either side of k on the home row and now stand in for the left/right arrows: l opens, j backs out.

Removes detailScrollOff, scrolledDetail, and scrollActive's detail case.

Two wrapping bugs the change exposed

Both were invisible in the alt screen and became real once rows were printed to the terminal:

1. lipgloss padding reached the scrollback. Style.Render pads a Width(n) block — after wrapping it runs alignTextHorizontal, which pads every row out to the widest line or to n. Printed rows kept that padding, and it came along when the text was selected and copied. wrapLines now calls ansi.Wrap, which is the wrap lipgloss itself runs (via x/cellbuf) before the align pass, so the padding is never added rather than added and trimmed back off. Measured on a representative result sentence at width 40: lipgloss padded 5/5 rows; ansi.Wrap, ansi.Wordwrap and cellbuf.Wrap padded 0/5 with identical line breaks. Boxes and modals keep Style.Render — their padding is structural.
2. Body text wrapped to the full terminal width. On a 200-column terminal a paragraph was set to 194 columns, which is hard to read back across and orphans the last words on their own line (193 + 7 for a typical result sentence). proseWidth caps the measure at 100 columns and keeps the existing floor; narrow terminals are unchanged.

x/ansi was already in the module graph as a lipgloss dependency; go.mod only promotes it to a direct require.

Verification

• go build ./..., go vet ./cmd/lk, go test ./cmd/lk pass.
• The clear must land after the alt screen is left — tea.Println is silently dropped while the alt screen is active, and a raw write inside a Cmd is not ordered against the event loop (Program.Send only queues the message), so the sequences ride along inside the print. Verified by driving a bubbletea program of the same shape through a pty and ordering the captured bytes: enterAlt @18 → exitAlt @68 → ESC[3J @89 → job text @93, and again on the second open/close cycle.
• Tests added cover logic only, not rendered output: the append/reprint decision (detailTail), the wrap measure arithmetic (proseWidth), that wrapLines rows carry no right padding, and the key handling (asserting on model fields).