test: cover ShellCommandRunner edge behavior

[dottie-weaver]

Adds 6 edge-case tests for ShellCommandRunner, complementing the V1 tests added in #30, closing #22.

Tests added

Test	What it covers
run_captures_both_stdout_and_stderr_simultaneously	Both streams captured independently
run_captures_stderr_alongside_nonzero_exit	stderr + non-zero exit together
run_silent_command_produces_empty_output	Silent commands yield empty strings
run_timeout_status_is_timed_out	Timeout produces TimedOut status
run_concurrent_calls_isolate_temp_files	4 parallel runs with isolated capture
launch_returns_immediately_without_blocking	launch() returns before process exits

Launch failure — deliberately untested

LaunchFailed / CommandLaunchError for a missing shell binary is documented but not tested live because:

1. shell_command() reads $SHELL at invocation time. Overriding requires env::set_var, which is unsafe in Rust 2024+ and unsound in multi-threaded tests.
2. No dependency-injection point exists for the shell path — shell_command is a private free function.

The existing shell_runner_reports_launch_failure test in #30 verified that nonexistent commands produce Exited(nonzero) (the shell itself exits), not LaunchFailed. That documents the actual behavior.

The four LaunchFailed code paths are structurally simple early-returns; full coverage would require a trait-based shell provider, out of scope for V1.

CI fix note

The initial push had run_timeout_preserves_partial_stdout which asserted partial output after kill. This is OS-dependent (Linux file buffering means data may not flush before kill()). Replaced with run_timeout_status_is_timed_out which tests the invariant that matters: the status is TimedOut.

Autonomy decision

Safe to auto-merge on approval — no workflow files changed, all CI-green locally.