deps(backend): bump litellm to 1.101.0

[MaxEriksson2000]

Changes

• litellm==1.98.0 → 1.101.0 in backend/pyproject.toml; uv.lock regenerated in the devcontainer (uv 0.12.2). The lock diff is the litellm version + hashes only — no transitive dependency changes.
• One new entry in backend/tests/warning_filters.py: pydantic 2.13 emits a UserWarning ("Item 'summary' on TypedDict class 'ChatCompletionReasoningItem' is using the ReadOnly qualifier…") the first time it builds litellm's Message schema in a process. Our warnings-as-errors policy turned that into a failure in test_litellm_api_key_override; the ignore carries a resolution path (bump once upstream drops ReadOnly from the TypedDicts pydantic consumes, or once pydantic stops warning). At runtime it is a single stderr line per process and has no functional effect.

Skips 1.99.x and 1.100.x deliberately: 1.100.0 shipped Python-compat regressions in response validation that 1.101.0 fixes (BerriAI/litellm#39399).

Why

Three minor releases behind. The change that matters for us is BerriAI/litellm#38593: temperature/top_p sent to gpt-5.5 / gpt-5.6 / gpt-6 are now dropped under drop_params=True instead of forwarded. On 1.98.0 they were forwarded, OpenAI answered 400 ("Unsupported value: 'temperature'…"), which matches none of the patterns in unsupported_param(), so users saw "The AI provider rejected the request". Verified with get_optional_params(..., drop_params=True) on both versions; gpt-5.4 keeps honoring the params as before.

Also relevant, verified against our integration surface (acompletion/aembedding/atranscription/aimage_*, model_cost, get_model_info, get_supported_openai_params, token_counter, litellm.exceptions):

• Both upstream breaking changes (prompt_token_calculator deleted, cerebras get_supported_openai_params) touch nothing we use.
• Bundled cost map grows 3040 → 3818 entries (claude-fable-5-1, gpt-6-astra, gemini-3.8-flash, deprecation_date on 49 mainstream models, sonnet-4-6 max_output_tokens 64k → 128k). Only matters for deployments without egress to GitHub — the map is otherwise fetched live at boot.
• Top-level anyOf/oneOf/allOf in tool schemas are flattened for OpenAI/Azure chat completions (#38839, #38870); previously a 400 for MCP tools with such schemas.
• token_counter differs by +2–3 tokens on OpenAI-tokenized paths, unchanged for Claude.
• The advertised lazy-import memory win (#39121) does not materialize: 194 → 208 MB RSS, 2299 → 2344 modules measured on the same host.

Follow-ups, out of scope here: gate the temperature slider on the new default_reasoning_effort map key (gpt-5.6 still advertises temperature in get_supported_openai_params but the value is now silently dropped), and use thinking_always_on to hide "none" in the reasoning-effort select.

Planning

Separate dependency-maintenance PR from the LiteLLM upgrade review; no tracking issue.

Testing

Inside the devcontainer on the worktree checkout:

• The six test files that patch litellm internals (test_image_generation, test_enrich_with_litellm_metadata, test_list_available_models, test_deprecation_lookup, test_deprecation_enrichment, test_litellm_api_key_override): 76 passed
• tests/unit + tests/unittests + tests/validation (-n 4): 4928 passed
• tests/integration/credentials/: 12 passed, 1 xfailed (same as develop)
• ruff check / ruff format --check clean; all pre-commit and pre-push hooks pass

Screenshots

n/a

[github-actions]

🧹 Dead-code & unused-dependency report

Advisory — never gates the PR. Whole-repo scan, so some findings may be false positives (dynamic dispatch, framework hooks, runtime-resolved imports). Triage before removing.

Scan	Tool	Findings
Unused Python code	vulture	1
Python dependencies	deptry	0
Frontend files / exports / deps	knip	0

Unused Python code — vulture (1)

src/eneo/mcp_servers/domain/entities/mcp_server.py:48: invalid syntax at "# type: the ask path mints a scoped token for it instead of sending stored"
src/eneo/image_models/domain/image_model.py:14: unused import 'ImageModelsDB' (90% confidence, 1 line)

Python dependencies — deptry: ✅ clean

Scanning 878 files...

Success! No dependency issues found.

Frontend — knip: ✅ clean

$ knip-bun --no-exit-code
No Svelte config file found in /home/runner/work/eneo/eneo/frontend - using SvelteKit's default configuration without an adapter.
No Svelte config file found in /home/runner/work/eneo/eneo/frontend - using SvelteKit's default configuration without an adapter.