Continue the whole-package mutation sweep, fortifying the command layer:

- init.py: pin _pick_template's stdin/stdout `or` (either stream piped -> usage
  error) and that error's exit_code; the missing-questionary exit_code; the
  agents-host derivation replacing only the FIRST "streaming" (replace count=1);
  the [:300] install-error truncation; and the DIRECTORY/--here conflict
  exit_code. Mark the don't-launch `False` on the install-failure branch
  `# pragma: no mutate` -- run_init exits on the failed step before consulting it.
- llm.py: assert `-o json` forces JSON output without the global --json flag.
- transcripts.py: assert an errored transcript surfaces its own error message,
  not the generic fallback.
- setup.py: assert subprocess capture_output, the 300s/120s skill timeouts, the
  install-hint command slice, the remove failure-detail fallback, _copy_tree's
  mkdir parents/exist_ok, and best-effort rmtree(ignore_errors). FakeRun now
  records per-call kwargs so timeouts/flags are assertable.

Tests only (plus one pragma); no behavior change.

- runner.py: pin find_free_port's bind-to-port-0 (OS-assigned ephemeral port),
  the port-range error message bounds and its exit_code, the 0.2s poll interval
  in wait_for_port, run_setup's success sentinel returncode, and the
  capture_output/check/text kwargs passed to subprocess.run.
- scaffold.py: assert the unknown-template / template-missing exit_codes, that a
  nested target's parent dirs are created (target.mkdir parents=True), and that a
  re-scaffold over an existing tree is tolerated (exist_ok). The two _copy_tree
  mkdir parents=True flags are marked `# pragma: no mutate` -- equivalent mutants,
  since the copy walk always creates a node's parent before the node.

Tests only (plus two pragmas); no behavior change.

- auth/loopback.py: assert the callback server thread is a daemon and that the
  cleanup join uses the bounded 5s timeout (spying threading.Thread while the
  server really serves, so shutdown() doesn't block).
- agent/audio.py: pin the DuplexAudio blocksize max(1, rate//10) floor for a tiny
  device rate.
- agent/session.py: pin the bounded 10s wait on ready_event in the send loop, and
  that the capture thread is a daemon.

Tests only; no behavior change.

- macos.py: assert the missing-swiftc / compile-failure exit_codes, the swiftc
  subprocess capture_output/text/check kwargs, the cache-dir mkdir parents (nested
  path) and exist_ok (pre-existing dirs), the _cleanup_process terminate guard +
  2s wait backstops + stderr-pipe close, the `returncode >= 0` boundary, and the
  chunk-frames = sample_rate//10 helper arg. The module-cache mkdir parents=True
  is `# pragma: no mutate` (equivalent: cache_dir is created the line before).
- render.py: assert a turn event missing end_of_turn reads as a partial (False).
- session.py: assert the parallel source workers are daemons; the 0.1s join poll
  interval is `# pragma: no mutate` (a responsiveness/CPU tradeoff, not behavior).

Tests only (plus two pragmas); no behavior change.

…-summary

Two follow-ups found by a full re-sweep after the rebase:

- setup.py _run: also assert the subprocess.run text=True and check=False kwargs
  (capture_output was already pinned) so all three are mutation-covered.
- config.py _validation_summary: mark exc.errors(include_url/include_input=False)
  `# pragma: no mutate` -- equivalent mutants, since the summary reads only loc+msg
  and never the url/input fields those flags toggle.

Tests only (plus one pragma); no behavior change.

Developer-experience follow-up. The mutation gate is diff-scoped, so auditing
existing code against its bar meant rebuilding a throwaway sweep script each time.
Promote it to scripts/mutation_sweep.py: it reuses the gate's own
collect/cover/survive engine over every line of the named files (or the whole
package), reports surviving mutants (exit 1) and an UNCOVERED bucket separately,
and is robust to the line-number shifts that make per-line checks brittle.

Document the workflow in AGENTS.md next to the diff-scoped gate, including the
reminder to pass `--timeout` to the coverage-refresh pytest run — the default
suite leaves per-test timeouts opt-in, so a deadlocked test otherwise wedges the
whole run instead of failing fast.