fix: detach spawned tsmd stdio so it can't hang an agent harness

[tashian]

The bug

The CLI auto-spawns tsmd and set cmd.Stderr = os.Stderr. Because tsmd is long-lived, it kept a copy of the caller's stderr fd open for its entire life. Under an agent harness like Claude Code's Bash tool — which considers a command finished only when its stdout/stderr pipes reach EOF — that fd never closes, so the first tsm call that has to spawn the daemon hangs the harness forever, even though the tsm process itself already exited.

In a normal terminal the same code is harmless (a tty isn't waiting for EOF), which is why it only bites under agents. This is a likely contributor to first-call credential-access friction for agents.

The fix (spawn() in internal/daemon/lifecycle.go)

• Redirect the daemon's stdout+stderr to a log file — new paths.DaemonLog() → <dataDir>/tsmd.log — falling back to /dev/null. Never the caller's stdio.
• Start it in its own session (Setsid) so signals to the CLI's process group don't take the daemon down, and Process.Release() to fully disown it.
• Drop the stdout socket-path readback. --socket is passed explicitly, so readiness comes from waitForSocket() polling the socket for liveness. (Bonus: removes a 10s-timeout failure path when the daemon is slow to print.)

Testing

• New deterministic TestSpawn_RedirectsDaemonStdioToLogAndDetaches: a fake python tsmd binds the socket and writes a marker to stderr; the test asserts the marker lands in the log file (proving stderr is redirected, not inherited) and that the daemon runs in its own session (getpgid(pid) == pid). Red before the fix (stderr leaked to the test's own stderr; 10s stdout-readback timeout), green after.
• Added paths.DaemonLog() tests (default + XDG_DATA_HOME).
• Full Go suite, gofmt, and go vet clean.
• Live: with no daemon running, a cold tsm status spawned the daemon and returned in ~1s (previously hung), daemon detached and alive, ~/.local/share/tsm/tsmd.log created.

Notes

• Independent of feat: gate confirm-mode tsm run on biometric presentability, not caller TTY #16 (that PR is the confirm/biometric gating; this is process lifecycle). No conflicts expected — disjoint files.
• Startup-failure detection shifts from "missing stdout line" to waitForSocket timing out, with the reason captured in tsmd.log.