#2369 feat(desktop): the call thread places each planning turn's work under the line that asked for it (call tab 2/3)

Second of the Call tab rollout, stacked on #2368. A pure model and its tests; nothing draws it until 3/3.

call-thread.ts builds the Call tab's one thread: the plan's calls one after another, each line said on them, and each planning turn's work right under the line that asked for it. It works the way a ChatGPT Voice call lands in its chat thread, in the coding agent tab's style.

• The link is the row id a turn names its ask by (asked, #2368) and a line names its row by (metadata.rowId).
• An ask spanning two lines stands under the later one.
• A turn no line on screen asked for (typed, lines not read yet, or past what the transcript holds) stands where it started: inside its call after the lines, between calls once its call had ended, or ahead of every call.
• A turn is placed in any shape (CallTurn<T>), as agreed with the one-agent-runtime rollout, so its later one-wire PR only swaps where the turns come from.
• A line heard on the live call now names its row in metadata the way a stored line does (transcript-model.ts), so the thread reads one key.

Tests (call-thread.test.ts, through the public export):

• Work stands under its line, Luke's answer follows it, and a two-line ask stands under the later line.
• Unlinked turns are placed by time, in all three places.
• A heard live line links like a stored one. This test fails without the heardRow change.

There's no Mac UI in this PR. ./scripts/verify.sh was not run (Linux VM). CI cannot verify the Mac; LUKE-159 is the manual Mac pass.

🤖 Generated with Claude Code

---

Open workspace in Conductor

---