simulate: tool output behind t, and stop reprints duplicating the job view

[u9g]

Two changes to the job detail view, together because the second is unusable without the first, plus a small --project fix that rode along.

Tool output moves behind t

Tool outputs are payloads written for the model, not for a reader — a policy blob or a lookup table's worth of JSON sitting between two spoken turns buries the conversation the job is about. They were previously always shown, clipped to 80 characters, which is the worst of both: still noisy, and too short to answer what a tool actually returned when a job failed on a tool call.

They are now off the transcript by default and shown whole on t. The hint reads t show tool output / t hide tool output, and is omitted when the open job has no tool output to show.

The tool call itself stays on the transcript either way — which tool ran, with a preview of its arguments, is part of reading what the agent did. Argument previews now clip on a rune boundary; args[:80] is a byte slice and these payloads carry guest names and quoted speech, so a rune landing on the boundary was emitted halved.

A reprint erases the copy it replaces

Toggling anything that rewrites the middle of the job body — which is what tool output does — used to leave a duplicate copy of the whole job in the scrollback, one per keypress. ctrl+L already had this: pressing it a few times stacked copies.

Since #928 the detail view is printed into the terminal rather than windowed in the alt screen. flushDetail diffs each render against detailPrinted and prints only the growth, so the common case (a log line arriving, the transcript landing with the summary) costs one appended tail. A render that is not an append can't be patched in place — scrollback is immutable — so detailTail falls back to returning the whole body. But clearScrollback was only prepended when detailPrinted == "", i.e. on the first print of a job, so every later whole-body reprint landed below the copy it was meant to replace.

The first condition widens to any reprint — nothing printed yet, or a render that isn't an append — and clearScrollback is prepended in both cases.

For ctrl+L the asymmetry is why it looked erratic: the Logs: block is the tail of renderDetail, so turning logs on was a pure append and behaved, while turning them off shortened the render and reprinted everything.

Notes

• clearScrollback is ESC[3J, so it takes the pre-TUI shell scrollback with it. That was already true of every job you opened (see the comment on the const); it is now also true per toggle. Reaching rows that have scrolled off isn't possible any other way.
• No tests. The reprint behavior is an ANSI prefix on a tea.Println, not something a detailTail unit test can observe, and asserting on rendered TUI output is what we avoid in this file. The transcript and hint logic were checked with throwaway tests — default hides output, t shows it whole, hint absent when there is nothing to show, rune-boundary clip — none of them committed.

Verification

go build ./..., go vet ./cmd/lk, gofmt, and the existing ./cmd/lk suite pass. The rendering and hint logic were exercised directly against a fixture shaped like a real emergency-dispatch job. The interactive path — actually pressing t and ctrl+L in a live TUI — has been spot-checked against a --view run but is worth a second pair of eyes.