Clean up PLC0415 noqa tech debt (inline imports)

[snopoke]

Goal

Remove the 303 # noqa: PLC0415 suppression comments added when the import-outside-top-level ruff rule was enabled. Each suppressed import should either be moved to module level (the common case) or kept as a lazy import with a documented reason (circular import avoidance or startup-time reduction per docs/agents/django_performance.md).

Context

PR #2993 enabled PLC0415 to prevent new inline imports, but suppressed all 308 existing violations with # noqa: PLC0415 as a baseline. The policy from AGENTS.md is clear:

Don't use local imports for any reason other than to avoid circular imports or as a means to reduce startup time (reserved for specific imports)

So for each noqa comment:

• If the import is inside a function/method purely for convenience → move it to the top of the file
• If it's avoiding a circular import → keep the noqa, add a brief comment explaining why
• If it's lazy-loading a heavy library (see docs/agents/django_performance.md) → keep the noqa, it's intentional

The heavy libraries that should stay lazy: langchain_google_vertexai, langchain_google_genai, langchain_anthropic, langchain_openai, boto3 (these are also enforced by TID253).

ALWAYS run the test before completing to surface any import issues.

Counts by app (total: 303):

• 68 service_providers
• 37 utils
• 35 pipelines
• 20 experiments
• 18 evaluations
• 15 human_annotations
• 14 chat
• 12 help
• 12 web
• 10 teams
• 8 trace, 7 events, 7 documents, 6 channels
• smaller apps: api, assistants, chatbots, data_migrations, users, dashboard, participants, custom_actions, files, sso, ocs_notifications, audit

Tasks

Task 1: service_providers (68 violations)

• Task 1

Clean up the 68 # noqa: PLC0415 comments in apps/service_providers/. Key files: llm_service/main.py, models.py, messaging_service.py, speech_service.py, llm_service/runnables.py, llm_service/utils.py. The lazy-loaded LLM provider imports in llm_service/main.py are intentional (keep with note); others should be moved to module level unless they're avoiding circular imports.

Task 2: utils + pipelines (72 violations)

• Task 2

Clean up 37 violations in apps/utils/ and 35 in apps/pipelines/. Key files: utils/deletion.py, utils/restricted_http.py, utils/tests/, pipelines/models.py, pipelines/repository.py.

Task 3: experiments + evaluations + human_annotations (53 violations)

• Task 3

Clean up 20 violations in apps/experiments/, 18 in apps/evaluations/, and 15 in apps/human_annotations/.

Task 4: chat + help + web + teams (48 violations)

• Task 4

Clean up 14 in apps/chat/, 12 in apps/help/, 12 in apps/web/, 10 in apps/teams/.

Task 5: Remaining apps (30 violations)

• Task 5

Clean up the remaining ~30 violations across trace, events, documents, channels, api, assistants, chatbots, data_migrations, users, dashboard, participants, custom_actions, files, sso, ocs_notifications, audit.

Learnings

• Task 2 (utils + pipelines): All remaining noqa annotations were already properly documented from previous work — no new changes needed. The remaining 41 annotations are all justified (circular imports, test setup requirements, Django LOGGING startup).
• Task 3 (experiments/evaluations/human_annotations): Most violations were convenience imports safely moved to top level. Circular imports were kept with descriptive comments. The langchain_core.messages import in experiments/tasks.py was kept lazy to avoid langchain import on startup.

[claude]

Task 2 & 3 complete

Task 2 (utils + pipelines): All remaining noqa annotations were already properly documented from prior work (PR #3024). No additional changes needed — the remaining 41 annotations are all justified circular imports, test setup requirements, or Django LOGGING startup issues.

Task 3 (experiments/evaluations/human_annotations): Cleaned up 53 violations in PR #3030.

• Moved convenience imports to module level in: experiments/models.py, experiments/views/experiment.py, experiments/management/commands/, evaluations/evaluators.py, evaluations/tasks.py, evaluations/views/, evaluations/tables.py, evaluations/tests/, human_annotations/tests/test_views.py
• Kept lazy with explanatory comments for: 14 circular imports in experiments/models.py, langchain startup import in experiments/tasks.py, circular imports in evaluations/models.py, evaluations/utils.py, and human_annotations/models.py

All 300 tests pass.