Feature/llm hardening core slim (#52)

[scrrlt]

Key changes
Client & factory:

• LLMClient.complete now forwards history as messages (prevents context drop).
• Factory preserves provider context for local routes in error mapping (so metrics/logs show provider).

Streaming & concurrency:

• Routing stream now enforces circuit breaker + TrafficController (bulkheads + rate token bucket).
• TrafficController guards against invalid RPM (requests_per_minute <= 0).
• Cohere streaming is rejected explicitly (LLMConfigError) until proper parser support is added.

Providers & SDKs:

• Anthropic provider: full integration (complete + stream), sanitized kwargs, stream handling.
• Local provider: local LLM server integration and streaming support.
• OpenAI provider: warn and disallow DISABLE_SSL_VERIFY in production (prevents accidental MITM).
• Base provider: treat LLMAPIError.status_code None as retriable/recordable.

Error mapping, telemetry, and utils:

• map_error improvements and tests.
  is_local_llm_server handles schemeless URLs and private IP detection.
• Circuit breaker logging improved and unexpected-state visibility added.
  CI & lint:
• .github/workflows/tests.yml: fail fast (removed "|| echo ..."), un-ignored tests/agents.
• pyproject.toml: re-enable B006 and stop ignoring B904 (catch mutable-default bugs and force exception chaining).
• pre-commit/ruff formatting applied and committed.

Touches files in LLM

ruff: passed (files touched only)
pre-commit: ran and auto-fixed formatting
mypy --strict: no errors for targeted modules
pytest tests/services/llm: 42 passed (core tests)
pytest tests/agents: agent tests pass (now included in CI)
bandit: ran across changed files (low-level findings limited to asserts in tests — expected)

Behavioral notes & compatibility

• Non‑breaking for most users — behavior improves reliability. Minor, intended behavior changes:
• Cohere streaming now raises LLMConfigError (explicit rejection) until supported.
• DISABLE_SSL_VERIFY in production now raises LLMConfigError (prevents unsafe production config).
  Developer-facing rule changes:
• B006 enabled prevents mutable default hazards.
• B904 removed from ignore enforces exception chaining where appropriate.

[pancacake]

Thanks bro!!!