Make the lint gates enforce small, single-screen functions (the friction a
coding agent hits most), then refactor the functions that exceeded the new
limits without changing behavior.

Gates:
- ruff: add C90 (mccabe) to select; set mccabe.max-complexity=10,
  pylint.max-statements=40, max-branches=12, max-returns=6, and lower
  max-args 10 -> 8. Narrow the commands/** ignore to [FBT001, FBT003,
  PLR0913] so command *bodies* are held to the same length/branch limits
  as the rest of the package (only the Typer-signature realities stay
  exempt).
- xenon: tighten --max-modules B -> A, so no module may average worse than
  grade A. (xenon measures branching only; raw length/arg limits live in
  ruff.)

Refactors (behavior-preserving):
- commands/stream.py: replace the ~250-line nested-closure command body with
  a module-level _SourceOptions, a _StreamSession class, and
  _validate_sources/_dispatch helpers; collapse the duplicated run/
  run_parallel lifecycle into one _guarded helper.
- commands/init.py, llm.py, agent.py: lift logic out of the nested command
  closures into module-level helpers.
- agent/session.py: bundle the static run config into a frozen AgentRunConfig
  (run_session 10 args -> 6) and extract _session_update_message/
  _receive_loop; name the untyped-SDK boundary with type aliases so Any
  stays in one place.
- timeparse.py, code_gen/stream.py: small readability extractions to bring
  their module-average complexity to grade A.

Verified: ruff, ruff-format, mypy, pyright (strict + tests), xenon,
vulture, deptry, import-linter, build/twine all pass; 903 tests pass at
99% coverage with 100% patch coverage.