fix: strip tool invocation artifacts from assistant text to prevent context explosion (#1492)

[xielixing]

Fix #1492: Strip tool invocation artifacts from assistant text to prevent context explosion

Problem

When a model outputs <tool_calls> XML as plain text content (instead of structured tool-call deltas), the system re-parses it as actual tool calls, causing infinite recursion and context explosion. The leaked XML passes through to the next round's context unchanged.

Root Cause

In round_executor.rs, after receiving the assistant's text content, the code calls strip_bitfun_memory_citations() but does NOT check for tool invocation artifacts. Leaked <tool_calls> XML passes through as-is to the next round's context, where it gets re-parsed and causes the infinite loop described in #1492.

Fix

The write_content_sanitizer.rs module already has reusable contains_tool_invocation_artifacts() and strip_tool_invocation_artifacts() functions — they were previously only applied to Write tool content arguments. This PR applies them to assistant text output as well:

1. round_executor.rs: After strip_bitfun_memory_citations, conditionally check and strip tool invocation artifacts from the assistant text, with a warn! log for observability.
2. write_content_sanitizer.rs: Added regression test strips_leaked_tool_calls_xml_from_assistant_text covering the exact [Bug]: LLM 工具调用 XML 泄露导致快速无限膨胀污染上下文 #1492 scenario.

Validation

• cargo check -p bitfun-core — passes
• cargo test -p bitfun-core write_content_sanitizer -- --nocapture — all 5 tests pass (4 original + 1 new regression test)

Files Changed

• src/crates/assembly/core/src/agentic/execution/round_executor.rs (9 lines added)
• src/crates/assembly/core/src/agentic/execution/write_content_sanitizer.rs (23 lines added)

Closes #1492