fix(sdk): nudge instead of finishing on a malformed-function-call content leak (gemini/OpenRouter)

[Viz-gov]

Problem

gemini-3.5-flash (and other Gemini models) intermittently hit Gemini's MALFORMED_FUNCTION_CALL finish reason — especially on tool-heavy / multimodal turns. When served via OpenRouter's OpenAI-compat layer, that native reason is remapped to a normal finish_reason="stop" (exposed only as native_finish_reason), and a scrap of the botched call leaks into the message content (e.g. "5", "}", "000Z").

classify_response then sees non-blank content + no tool call → CONTENT → _handle_content_response, which sets execution_status = FINISHED. The episode ends on what is really a fumbled tool call. Empty / reasoning-only turns are already recovered by _send_corrective_nudge, but a non-empty stray token slips straight through and silently kills the run.

Fix

In _handle_content_response, if the only visible content is an implausibly-short fragment (≤12 chars, single line) with no tool call, treat it as a malformed-function-call leak and send the existing corrective nudge instead of finishing. Substantive text answers are unaffected.

Evidence

Task auth-service-latency-investigation (multi-app-200, gemini-3.5-flash via OpenRouter, OpenHands SDK):

	reward	steps
before (3/3 prior runs)	0.0	dead at 4 (bare "5")
with this patch	0.625	278

The leak still occurs ~30× in the run — every one is now caught and nudged, and the agent recovers and completes the task. Canyon (text-only, no images) never tripped it, consistent with the multimodal trigger.

Notes / follow-ups

• A more principled fix is to branch on native_finish_reason == "MALFORMED_FUNCTION_CALL" for the OpenRouter path, or use the native Google provider where finishReason=9 is visible and retryable. That signal isn't currently plumbed through; this content-heuristic backstop is provider-agnostic and complements re-enabling stuck detection.
• Refs: Gemini MALFORMED_FUNCTION_CALL = finishReason 9; OpenRouter remaps to stop (pydantic-ai chore(ci): use GITHUB_TOKEN for label and stale workflows OpenHands/software-agent-sdk#2844, agno chore(deps): bump lewagon/wait-on-check-action from 1.7.0 to 1.8.0 OpenHands/software-agent-sdk#3649).