fix: apply empty-content guard to empty-string chat tool outputs

[ioleksiuk]

Summary

Extends the fix in #3312 (`fix: avoid empty chat tool outputs`) to one case the original guard missed.

#3312 added an empty-content guard that swaps the tool message body with `_OMITTED_TOOL_OUTPUT_PLACEHOLDER` when the converted content is empty. The guard was placed inside the `else` branch of `isinstance(all_output_content, str)` at `chatcmpl_converter.py:749-769`, so it only triggered when the output was a non-text-only list. When the output is the empty string `""`, `extract_all_content("")` returns `""` (the string branch), `tool_result_content` is set to `""` without ever reaching the guard, and the Chat Completions message ends up with `"content": ""` — exactly the broken shape #3310/#3312 aimed to prevent.

Repro on current `main` (before this PR):

```python

from agents.models.chatcmpl_converter import Converter
items = [
... {"type": "function_call", "id": "fc_1", "call_id": "c1",
... "name": "get_status", "arguments": "{}"},
... {"type": "function_call_output", "call_id": "c1", "output": ""},
... ]
[m for m in Converter.items_to_messages(items) if m.get("role") == "tool"]
[{'role': 'tool', 'tool_call_id': 'c1', 'content': ''}]

Chat Completions API rejects: messages.X.content must be non-empty

```

Lift the guard above the `if/else` so both branches go through it. `not tool_result_content` is truthy for both `""` and `[]`, so the single check covers both empty shapes.

Test plan

• New `test_items_to_messages_with_empty_string_function_output_uses_placeholder_by_default` — verifies the default-mode placeholder + warning for `output=""`.
• New `test_items_to_messages_with_empty_string_function_output_raises_in_strict_mode` — verifies strict-mode UserError.
• `uv run pytest tests/models/test_openai_chatcompletions_converter.py -k empty` — 4 passed (both pre-existing empty-list tests + the two new empty-string tests).
• `uv run ruff check` on the touched files — All checks passed.

Issue number

N/A — found while auditing for the same fix pattern that #3312 applied to empty lists.

Checks

• I've added new tests (if relevant)
• I've added/updated the relevant documentation (no doc changes needed)
• I've run `make lint` and `make format`
• I've made sure tests pass

[ioleksiuk]

Closing — on second read of #3310, the issue reporter explicitly listed "String tool outputs remain unchanged." as a deliberate edge case. The placement of #3312's guard in the else branch was intentional, not an oversight.

Concrete reasons this PR doesn't make sense after all:

1. ChatCompletionToolMessageParam.content is Required[Union[str, Iterable[...]]] per openai-python — empty string is a valid value of str, the API type does not reject it.
2. When a user writes return "" from a function tool, the empty string is their intentional output. The SDK should not silently rewrite it to [tool output omitted].
3. The bug fix: #3310 avoid empty chat tool outputs #3312 fixed was lossy conversion: a tool returns image/audio content → the converter strips it → message ends up empty. Empty-string output isn't lossy — it's faithful to the tool's return value.

My PR conflated "empty content after filtering non-representable parts" with "user explicitly returned empty string" — those are semantically different.

Apologies for the noise.