fix(cli): record the command in plain simulate's run manifest

[isarasua]

What changed and why

A run manifest records how the run was launched (source.kind, source.invocation). Plain
usersim simulate passed source_kind="cli" to write_simulator_manifest. The manifest fills in
the invocation only when it detects the kind itself, so plain runs always recorded an empty
source.invocation.

For example, usersim simulate --locale en_US --num-rows 50 --max-turns 6 recorded:

"source": {"kind": "cli", "invocation": ""}

So a month later, the manifest couldn't say which flags made the run.

The override is gone. The manifest now detects the source itself, as it already does for
simulate --inputs, eval and the other writers. A CLI run records its command line (sys.argv).

As agreed in the review of #30.

How it was verified

• The test. test_simulate_helper_mints_new_run_when_skip_existing_false now fixes sys.argv
  and turns off SLURM and notebook detection. It asserts source.kind == "cli" and the exact
  invocation. It failed before the change (empty invocation) and passes after.
• Codex found nothing in the change. Its one nit, isolating notebook detection in the test, is
  fixed.
• Checks. make check-all passes, and the full suite passes: 3,835 tests.
• Readers. Nothing in src, the docs or the notebooks reads source.kind or
  source.invocation. Only tests do.
• Merges cleanly with feat(cli): run stored episode rows with simulate --inputs #30, feat: opt-in replay of the assistant's own earlier reasoning #31 and fix(storage): write simulation runs under a lock and publish them whole #34.

Note

Under SLURM, or when _simulate_to_parquet is called inside a notebook, the manifest now says
slurm (with the job id) or notebook, as the other writers' manifests already do. A command-line
run outside SLURM records cli and its command.

🤖 Generated with Claude Code

[3mei]

Ready to merge!