Isolate classifier client/config changes from import refactor

[sheepdestroyer]

Description

CodeRabbit flagged that PR #272/#273 (sys.path import cleanup) mixed two
unrelated concerns:

1. Import refactor: removing sys.path.insert hacks, adding
   __init__.py, normalizing imports — the stated objective of the PR.

2. Classifier client rework: adding get_classifier_client() with
   dedicated httpx client, moving api_base from hardcoded URL to
   os.environ/LLAMA_CLASSIFIER_URL, adding classifier env var to
   .env and config.

Impact

The classifier changes were functional and needed, but they landed in what
was labeled as a chore/refactor PR. This:

• Makes git blame and history harder to follow
• Increases review surface (reviewers must assess import cleanup AND a
  new HTTP client pattern simultaneously)
• Risk: if the classifier client had a bug, bisecting would point at the
  import cleanup PR, confusing diagnosis

Suggested Action

Since the PR already merged (#275), this issue tracks the retrospective
concern. Going forward:

• Keep PRs scoped to one concern
• New infrastructure (HTTP clients, config changes) should be its own PR
• Refactors should be pure refactors unless the functional change is a
  direct prerequisite

Related

• PR chore: eliminate sys.path.insert hacks — migrate to package-relative imports #275 (merged — contains both concerns)
• Issue Ollama routing broken: classifier llama-server (port 8080) not deployed — all requests fall through to ollama-deepseek-v4-pro #274 (Ollama routing / classifier deployment)