fix(agent-runtime): report subagent output token truncation as failure

[Totopo27]

Summary

Ensures subagent runs whose response hits the model output token limit (stopReason: "length" / "max_tokens") settle as failed with error code SUBAGENT_OUTPUT_TRUNCATED rather than falsely reporting completed.

Motivation & Root Cause

Resolves issue #1190.

In packages/agent-runtime/src/subagent.ts, SubagentRun.run() evaluated completion purely on whether lastReportText was non-empty:

if (!this.lastReportText.trim()) {
  return this.result("failed", "", { code: "SUBAGENT_NO_REPORT", ... });
}
return this.result("completed", this.lastReportText);

When a subagent's response exceeded the model's output token limit without issuing tool calls, the stream terminated with stopReason: "length". Because the message contained partial report text, the subagent silently settled as completed, handing the parent agent an incomplete report cut off mid-generation with no indication of truncation.

Key Changes

• Error Taxonomy (packages/shared/src/errors.ts):
  • Added SUBAGENT_OUTPUT_TRUNCATED to ErrorCodes.
• Subagent Runtime (packages/agent-runtime/src/subagent.ts):
  • Added outputTruncated?: boolean to SubagentRunResult.
  • Tracked lastReportTruncated on message_end when stopReason === "length" || stopReason === "max_tokens".
  • Settle run() as failed with SUBAGENT_OUTPUT_TRUNCATED whenever lastReportTruncated is true, surfacing the explanatory failure message while preserving the partial output under Its last output was:.
• Unit Tests (packages/agent-runtime/src/subagent.test.ts):
  • Added test verifying that a subagent ending on a truncated message reports status: "failed", code: "SUBAGENT_OUTPUT_TRUNCATED", and outputTruncated: true.
  • Added test verifying that a subsequent turn completing with stopReason: "stop" clears the truncated state and completes cleanly.

Verification

• Clean syntax and type verification across modified modules (errors.ts, subagent.ts, subagent.test.ts).
• Regression suite passing without functional regressions.
• Strict compliance with R6.1 (fix: contribution addressing verified upstream issue [Bug] 被输出上限截断的子智能体仍被报成 completed #1190).