fix(llm): silence litellm 'Provider List' stderr banner

[rolandpg]

Summary

LiteLLM's get_llm_provider helper writes a coloured Provider List: https://docs.litellm.ai/docs/providers banner to stderr via raw print() whenever it cannot resolve a provider from a model name (litellm/litellm_core_utils/get_llm_provider_logic.py:466). The banner bypasses Python logging entirely, so it leaks past ZettelForge's structlog setup.

This affects every operator running recall() — the background LLM-NER enrichment fires litellm via entity_indexer.extract_llm even when the user is only doing pip-only remember() / recall() without a working LLM. The first user trying the README's 30-second hello world (R1, just merged) sees ~40 of these banners during a single recall — exactly the wrong first impression now that the no-Ollama path is being marketed.

Reproduction (before this PR)

$ python examples/quickstart.py 2>&1 | head -20
Stored: note_20260427_222417_0d23 (created)
Entities: ['cve-2024-3094', 'apt28', 'cobalt-strike', 't1021']
[1;31mProvider List: https://docs.litellm.ai/docs/providers[0m
[1;31mProvider List: https://docs.litellm.ai/docs/providers[0m
... (~40 more) ...
Found: APT28 (Fancy Bear) targets NATO defense contractors with spear-phishing...

Fix

Route the lazy litellm import through a new _get_litellm() helper that sets litellm.suppress_debug_info = True after import. suppress_debug_info is the documented escape hatch for exactly this banner — it gates only the print statement, not exception raising, so:

• Real errors continue to propagate via litellm.exceptions.*.
• The existing llm_call_exception structured-log event still fires with the same fields.
• No behaviour change for callers with valid model names + API keys.

After this PR:

$ python examples/quickstart.py 2>&1 | grep -c "Provider List"
0

Verified end-to-end with a fd-2-capturing test harness — zero Provider List bytes leak through.

Tests

• 12 LiteLLMProvider tests pass (was 11 + new test_generate_silences_litellm_debug_banner).
• The new test mocks the litellm module, asserts the suppress flag flips True after a generate() call, and would catch any future regression that drops the helper or its side effect.

Out of scope

• The env-dependent test_import_error_raised_when_sdk_missing already failed in litellm-installed environments before this change. CI runs without litellm, so it passes there. Out of scope for this PR.
• Other litellm chatter (DEBUG-level LiteLLM Logging Details, etc.) goes through Python logging and is already managed by structlog level config — not in scope.

🤖 Generated with Claude Code

[Copilot]

Pull request overview

This PR reduces operator-facing stderr noise by silencing LiteLLM’s raw print() “Provider List” banner when provider resolution fails, ensuring it doesn’t leak past ZettelForge’s structlog setup during recall().

Changes:

• Added a _get_litellm() import helper that sets litellm.suppress_debug_info = True before calls to litellm.completion().
• Updated LiteLLMProvider.generate() to use _get_litellm() instead of importing LiteLLM inline.
• Added a regression test asserting the suppress flag is flipped on generate().

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.

File	Description
src/zettelforge/llm_providers/litellm_provider.py	Routes LiteLLM import through _get_litellm() and sets suppress_debug_info to silence the stderr banner.
tests/test_llm_providers.py	Adds regression coverage to ensure generate() enables litellm.suppress_debug_info.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 862f4f8846

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Guard suppress_debug_info mutation with a lock

litellm.suppress_debug_info is a process-global module flag, but this code snapshots and restores it without synchronization; with concurrent generate() calls, one thread can restore an outdated value while another call is still in litellm.completion(). In this repo that is realistic because providers are shared singletons (src/zettelforge/llm_providers/registry.py notes singleton instances) and LLM work runs on a background enrichment thread (src/zettelforge/memory_manager.py), so this race can both re-enable the stderr banner mid-call and leave the global flag in the wrong final state after both calls complete.

Useful? React with 👍 / 👎.