fix(coderd/x/chatd): keep task attempts alive while execute tools make progress - #27103

This PR was authored by Mux, an AI coding agent, working on behalf of Mike.

Problem

Chatd task retries respawn execute tool processes, leaving orphaned duplicates in workspaces (CODAGT-757). The earlier PRs in this stack make retries re-attach to the original process instead of starting a new one, but every task attempt still dies at a fixed 15 minute timeout. A legitimately long-running execute (the clamp allows up to 4 hours) can never finish within one attempt: the attempt times out mid-wait, the task retries, re-attaches, and times out again, burning retries on a command that is making progress.

What this PR changes

• Redefines the 15 minute task attempt timeout in coderd/x/chatd/tasks.go as a resettable idle window, and adds a non-resettable 24 hour absolute cap (maxTaskAttemptDuration). Both cancel the attempt with the existing retryable errTaskTimeout.
• Adds chattool.WithAttemptKeepalive / chattool.KickAttemptKeepalive (new coderd/x/chatd/chattool/keepalive.go) using an unexported typed context key. Kicking is a safe no-op when the keepalive is absent or the attempt already ended. taskAttemptContext installs the kick into the attempt context.
• Kicks the watchdog in chattool only, after each successful agent round-trip: StartProcess returns (foreground and background), every successful ProcessOutput round inside waitForProcess (including rounds where the process is still running), the successful re-attach snapshot, and the process_output tool. Other workspace tools finish well within the idle window and deliberately do not extend it.
• Documents the attempt watchdog and the keepalive scoping in coderd/x/chatd/ARCHITECTURE.md.
• Clamps wait_agent's model-supplied timeout_seconds to 15 minutes per call (maxSubagentWaitTimeout), so a sibling tool cannot ride a progressing execute's keepalives for hours on one blocking call; a wait_agent timeout stays a resumable non-failure. A single computer wait action is likewise clamped to 30 seconds.
• Extracts the settle-once watchdog timer that chatloop's streamSilenceGuard already used into coderd/x/chatd/internal/watchdog and backs both the stream-silence guard and the new idle window with it, so the concurrency-sensitive timer lifecycle exists once. Quartz timer tags are unchanged; the disarm-versus-fire race tests move to the shared package.

What this PR does not change

• Streaming code never kicks the watchdog. The 10 minute stream-silence guard and the effective 15 minute window for non-tool work are unchanged, so silent provider streams keep failing through chat-specific retry handling first.
• No database or API changes.

Test plan

• go test ./coderd/x/chatd/... (full chatd tree, passes)
• TestTaskAttemptContext_* tests with quartz.NewMock: kicks extend the attempt past the idle window, the 24h cap fires regardless of kicks, a no-kick attempt still dies at 15m, kicking after cancel or timeout is a no-op.
• Chattool tests: keepalive kicked once per successful start, poll round, and re-attach snapshot; process_output tool kicks on success.
• internal/watchdog tests: fires with its cause, reset restarts the window, the disarm-versus-fire race cancels at most once, disarm preserves the context cause.
• make fmt, make lint, make gen: clean, no dirty diff.

Stack

PR 3 of 3 for CODAGT-757: builds on #27102 (agent idempotency token), which builds on #27100 (execution ledger).