fix(renderer-client): normalize empty content like token client

[eligotts]

Summary

• Stacked on top of #<feat/renderer-inference-v1-generate PR>. Mirrors a coercion already present in OpenAIChatCompletionsTokenClient.normalize_for_comparison: tool-call-only assistant messages can be serialized with content=None or content="" depending on the upstream pipeline (reasoning parsers strip text to "", other paths leave None), and incremental-prompt prefix matching must treat them as equivalent or it spuriously falls back to MITO.
• renderer_client._normalize_for_comparison already drops None values from mappings, so a content=None message normalizes to "no content key" while content="" would stay as "content": "". We extend the same filter to also drop content == "", unifying the two shapes.

Test plan

• Existing renderer-client tests still pass.
• Manually verify a multi-turn rollout where one trajectory step has assistant content=None and the next prompt presents the same step with content="" (e.g. via a reasoning parser): bridge succeeds instead of falling back.

🤖 Generated with Claude Code

---

Note

Low Risk
Low risk: a small normalization tweak that only affects incremental prompt prefix-matching by treating content:"" like missing/None, reducing spurious bridge failures without changing generation behavior.

Overview
Improves incremental prompt prefix-matching in renderer_client._normalize_for_comparison by dropping mapping keys where content == "", making tool-call-only assistant messages serialized as content=None vs content="" compare equal.

This prevents format-only drift in upstream pipelines from forcing a full re-render/fallback when stitching multi-turn prompts.

^{Reviewed by Cursor Bugbot for commit 2aaac1a. Bugbot is set up for automated code reviews on this repo. Configure here.}