perf: cut TUI startup from ~20s to ~0.15s (lazy-import litellm + IPv4-first)

[gilsonolegario]

Summary

graybox.cli startup currently takes ~20s because importing it pulls in litellm, whose own import does a blocking network fetch of its model-cost map (and hangs ~20s on networks without routable IPv6). This PR makes importing the CLI ~0.14s — a ~150x speedup.

Changes

Two commits:

1. perf: lazy-import LLM modules for fast startup — litellm and its ~900 modules are no longer imported at module level. They load only when a command actually needs the LLM (organize/ask/chat/dupes/merge/edit/delete/refresh).
2. perf: prefer IPv4 and local cost map in litellm startup — two fixes applied before litellm is imported:
   • Reorder getaddrinfo results so IPv4 is tried first (avoids the ~20s hang on broken-IPv6 networks).
   • Set LITELLM_LOCAL_MODEL_COST_MAP so litellm's bundled local cost map is used instead of a network fetch.

Verification

• python3 -c "import time; t0=time.time(); import graybox.cli; print(time.time()-t0)"
  • before: 21.5s
  • after: 0.13s
• 245 tests pass.

Why

The interactive TUI (graybox.cli with no args) is the primary way people run Gray Box. Waiting ~20s for a splash screen that doesn't even use the LLM is a poor first impression and unnecessary work.

[Aaryanverma]

Hi @gilsonolegario,

Request changes:

• The lazy-import optimization is good and should be retained. However, graybox.ai.ai_service globally replaces socket.getaddrinfo, which changes DNS resolution behavior for the entire host process, not just LiteLLM. Please avoid the global monkey-patch if possible. First verify whether LITELLM_LOCAL_MODEL_COST_MAP=true eliminates the import-time network dependency by itself; otherwise isolate the IPv4 preference to the LiteLLM/networking layer.

• Also add a regression test ensuring importing/starting the CLI does not load LiteLLM.

[gilsonolegario]

Thanks for the thorough review — the points are valid and both are addressed in the latest commit.

• I verified empirically that LITELLM_LOCAL_MODEL_COST_MAP=true alone eliminates the import-time network dependency: with it set, import litellm performs zero getaddrinfo/network calls (litellm 1.93.0). The global socket.getaddrinfo monkey-patch has been removed entirely; the env var is still set right before the litellm import, which is all that's needed.
• Added tests/test_startup.py: a subprocess-based regression test asserting that importing graybox.cli and starting the CLI (main(['--help'])) does not load litellm.

Full suite passes (243 tests on this branch). Happy to adjust if anything else stands out.