Request content: keep the live answer until the stored one has landed

[fylorn]

Problem

In the request drawer's Content tab, a streamed answer types out live as expected, but right after the request finishes the answer turns into "not saved" and goes blank. Reopening the request shows the full answer.

Cause

Core records the request row and the bodies on separate threads (tw_store::task). The row and request_priced land first. The request body was stored when the request started, while the heads and the answer are written only after the request ends and has been redacted (Ending::settle → bodies channel → for_disk → bodies thread).

landed(d) accepted the stored detail as soon as any of heads / request body / response body was present, so the request body alone was enough. The drawer switched to a detail with no response yet and nothing fetched it again.

A second path made it stick: the live subscription was dropped as soon as the request left in-flight. request_priced can arrive before the last chunk and end, so the subscription stopped before end, and the retry loop that fetches the stored copy (settle) never ran.

Fix

• caughtUp(d, live) replaces landed(d): the stored detail counts only once it has every head and body the live content showed: heads, client request, the serving hop's request when it was sent separately, the client answer, and the upstream answer when it differs from what the client got (until the client copy lands, response_body is the upstream one). Without live content the old rule applies.
• useLiveContent keeps the subscription until it closes on its own (end, error, or not running) instead of stopping when the request leaves in-flight; it still unsubscribes when the drawer closes or switches request. onClosed gets the live content so settle can keep polling until the stored copy has caught up.

Verification

• pnpm typecheck, pnpm vitest run (1206 tests pass), with 5 new cases for caughtUp in dialect.test.ts.
• Reproduced with a mock that follows core's order at the end (row + request_priced first, last chunk + end 250 ms later, heads + answer on disk 2 s later). Before the fix the answer section went to "not saved" at the end of the stream; after it the answer stays and the drawer switches to the stored detail once it has landed. The default order (end before request_priced) is unchanged.

🤖 Generated with Claude Code