fix(gemini): stop splitting one tool call into two when thoughtSignature arrives late

[Ricardo-M-L]

Summary

`translate_stream_event()` in `agent/gemini_native_adapter.py` keys streaming tool-call slots on `(part_index, name, thought_signature)`:

```python

before

thought_signature = part.get("thoughtSignature") if isinstance(part.get("thoughtSignature"), str) else ""
call_key = json.dumps(
{
"part_index": part_index,
"name": name,
"thought_signature": thought_signature, # ← splits one call into two
},
sort_keys=True,
)
```

Because `thought_signature` is part of the dedup key, a single tool call whose chunks carry the signature inconsistently is split into two separate slots:

• slot 0: built from the early chunks → no signature, partial args
• slot 1: built from a later chunk → has signature, fuller args

Both slots are emitted as deltas, so the agent records two tool calls for what was logically one. On the next turn the slot without a signature is replayed back to Gemini, which then 400s with:

```
Function call is missing a thought_signature in functionCall parts.
```

How this bug surfaces in the wild

Reported by a user running `gemini-3.1-flash-lite-preview` through Hermes Desktop:

```
HTTP 400: Gemini HTTP 400 (INVALID_ARGUMENT): Function call is missing a
thought_signature in functionCall parts. ... Additional data, function call
`default_api:terminal` , position 2.
```

Position 2 == the third part of the assistant content. With the bug, after one logical `terminal` tool call the model's content has:

• part 0: text (assistant prose)
• part 1: functionCall (slot 0 — partial args, no signature)
• part 2: functionCall (slot 1 — full args, has signature)

Gemini sees the part-1 functionCall without a signature and rejects.

Fix

Dedup on `(part_index, name)` only. The signature is still surfaced through the per-chunk `extra_content` field, and the downstream streaming accumulator in `run_agent.py` already does latest-non-None-wins on `extra_content` per slot:

```python

run_agent.py — existing behavior, unchanged

extra = getattr(tc_delta, "extra_content", None)
...
if extra is not None:
if hasattr(extra, "model_dump"):
extra = extra.model_dump()
entry["extra_content"] = extra # late-arriving signature wins, single slot
```

So whichever chunk carries the signature gets it merged into the single slot.

Test plan

Adds a regression test:

```python
def test_stream_event_translation_does_not_split_slot_when_signature_arrives_late():
...
# Chunk 1 — no signature yet.
event_a = {"candidates": [{"content": {"parts": [
{"functionCall": {"name": "terminal", "args": {}}}
]}}]}
# Chunk 2 — same tool call, signature now present.
event_b = {"candidates": [{"content": {"parts": [{
"functionCall": {"name": "terminal", "args": {"cmd": "ls"}},
"thoughtSignature": "sig-late",
}]}}]}
...
# Both chunks must address the SAME slot.
assert a_tool.index == b_tool.index == 0
assert a_tool.id == b_tool.id
# The chunk that carried the signature surfaces it via extra_content.
assert b_tool.extra_content == {"google": {"thought_signature": "sig-late"}}
```

• Fails on main with `assert 0 == 1` (slot 1 created with the late signature, slot 0 left signature-less).
• Passes here.
• All 12 tests in `tests/agent/test_gemini_native_adapter.py` continue to pass.

🤖 Generated with Claude Code

[Ricardo-M-L]

Closing in favor of #28438, which is the same fix rebased onto current main (this branch was 2K+ commits behind and would conflict). #28438 has the same one-key dedup change plus a regression test that fails on main and passes with the fix.