fix(agents): preserve text streamed before tool calls in output_key

[vipin-v-nair]

Fixes #5590

Problem

When an LlmAgent is configured with output_key and runs in
StreamingMode.SSE while making tool calls, only the model text on the
very last tool-free event is written to state[output_key]. Any text
the model emits before or between tool calls is silently dropped, even
though it streams to the user. Reporters measure 60–70% retention loss
in production. A live reproduction against gemini-2.5-pro on Vertex
shows the intro and progress narration discarded, with only the
conclusion text reaching output_key.

Root cause

LlmAgent.__maybe_save_output_to_state only writes
state_delta[output_key] on events where Event.is_final_response()
returns True. That gate returns False for any event carrying a
function_call or function_response part. Under streaming, Gemini
emits non-partial events that bundle text with a function_call —
those events fail the gate, so their text is skipped. The post-
aggregation events that follow the tool calls are also rejected (they
contain function_response parts). Only the final tool-free event
passes, and state_delta[output_key] gets overwritten with just that
text. Every prior text segment from the same model turn is silently
discarded.

Fix

Add a new private helper LlmAgent.__maybe_accumulate_streaming_output
that receives the current event and a running accumulator string. It
returns the new accumulator value, and on applicable events also writes
the running value to state_delta[output_key]. The helper is invoked
from the call sites in _run_async_impl and _run_live_impl,
alongside (and after) the existing __maybe_save_output_to_state call.
Each call site holds a local output_accumulator: str that lives for
the duration of the invocation — no new instance state on the agent.

__maybe_save_output_to_state is byte-identical to main. The schema
path is preserved unchanged: when output_schema is set, the helper
short-circuits and the existing single-final-document validation path
runs as before. The function-response-only callback path is also
preserved (the helper short-circuits on empty text).

Tests added

• tests/unittests/agents/test_llm_agent_streaming_output.py::test_run_async_accumulates_text_around_tool_cal ls
  Regression test for Text accumulation issue for output_key #5590. Drives _run_async_impl with a stubbed
  _llm_flow that yields the canned event sequence the streaming
  flow produces (interleaved partial text, text+function_call,
  function_response, more text, final tool-free text), merges each
  event's state_delta into session state via
  session_service.append_event, and asserts on
  session.state["final_output"] — the value named in the bug
  report. Fails on main with
  AssertionError: 'Conclusion one. Conclusion two.' == 'Intro one. Intro two.Progress.Conclusion one. Conclusion two.';
  passes with this fix.

Verification

• pre-commit run --all-files — clean
• pyink --check --diff --config pyproject.toml src/ tests/ — clean
• isort --check src/ tests/ — clean
• Mypy New Error Check (full-repo two-branch diff) — 0 new errors,
  0 removed
• pylint clean on touched files (no new warnings; pre-existing
  warnings unchanged)
• pytest tests/unittests --ignore=tests/unittests/artifacts/test_artifact_service.py --ignore=tests/unittests/tools/google_api_tool/test_googleapi_to_openapi_converter.py
  — 5608 passed, 2266 warnings, 6 subtests passed in 436.97s
• Live reproduction against gemini-2.5-pro on Vertex confirms all
  three text sections (intro + progress + conclusion) now reach
  state[output_key]; pre-fix run had only the conclusion.

Design notes

I considered an alternative that adds a PrivateAttr accumulator to
LlmAgent and keeps the fix inside __maybe_save_output_to_state. I
went with the call-site approach because:

• It avoids adding state to the agent (no new instance fields).
• It keeps LlmAgent's field set unchanged.
• A PrivateAttr on LlmAgent would be the first in the agent class
  hierarchy — this bug didn't feel like the right place to introduce
  that pattern, even though PrivateAttr is already used elsewhere
  in the codebase (agents/invocation_context.py:221,
  sessions/session.py:53).

The trade-off: the regression test constructs an InvocationContext
directly, so it's mildly brittle to future changes in
InvocationContext's required fields. The 14 existing tests in
tests/unittests/agents/test_llm_agent_output_save.py are byte-
identical to main and continue to pass unchanged.

Out of scope

• Other tests in test_llm_agent_output_save.py use name-mangled
  access to internal state (agent._LlmAgent__maybe_save_output_to_state(event)).
  The new test deliberately asserts at a higher level
  (_run_async_impl + session.state[output_key]), but the existing
  tests are untouched.