chore(deps): resolve remaining 18 Dependabot security alerts in worker_plan

[neoneye]

Summary

Resolves the 18 open Dependabot security alerts in worker_plan/pyproject.toml (the 19th, mcp in mcp_cloud, was fixed by merging dependabot PR #816):

Package	Change	Alerts resolved
pillow	12.2.0 → 12.3.0	13 (heap OOB writes, decompression bombs, DoS)
soupsieve	2.8.3 → 2.8.4	2 (ReDoS, memory exhaustion)
json_repair	0.58.7 → 0.60.1	1 (circular $ref CPU DoS)
transformers	removed	2 (RCE in model loading paths)

Why transformers is removed instead of bumped

The only patched version is 5.5.0, a major-version jump requiring huggingface-hub >= 1.5. But transformers was never used by PlanExe code — it was only a hard dependency of llama-index-llms-openai-like < 0.6, used solely when a string tokenizer name is passed to OpenAILike (PlanExe never passes one). From 0.6.0 the package makes transformers an optional extra with a lazy import.

So instead:

• llama-index-llms-openai-like 0.5.3 → 0.7.0 (satisfies existing pins: llama-index-core 0.14.12, llama-index-llms-openai 0.6.12)
• llama-index-llms-openrouter 0.5.0 → 0.5.1 (its older pin required openai-like < 0.6)
• Removed transformers and its now-orphaned transitive deps tokenizers, safetensors, huggingface-hub (nothing else in the freeze depends on them)

Verification

• Fresh venv, replicated the CI install (worker_plan + mcp_cloud + worker_plan_database + frontend_multi_user): no resolver conflicts introduced (the fastapi/requests/uvicorn cross-manifest warnings are pre-existing on main)
• python test.py: all worker_plan tests pass; only failures are the mcp_cloud tests that need PostgreSQL (expected outside Docker/CI)
• Confirmed OpenAILike, ThinkingAwareOpenAILike and the openrouter module import and instantiate fine with transformers absent

🤖 Generated with Claude Code