Fix Qwen3-Coder tool parser and harden server against mid-stream client disconnects

[i1rr]

Two related server-resilience fixes triggered by an agent flow against mlx-community/Qwen3-Coder-30B-A3B-Instruct-4bit via mlx_lm.server.

1. Qwen3-Coder tool parser — nested <parameter=...> substrings in values

Problem

_parameter_regex used <parameter=(.*?)</parameter> with non-greedy .*?. When a value contained text like:

<parameter=file_text>
def example():
    # see <parameter=path>...</parameter> in the docs
</parameter>

the regex stopped at the first inner </parameter>, splitting the value mid-content. The truncated tail was then no longer valid JSON or a valid Python literal, so _convert_param_value raised SyntaxError from ast.literal_eval. That SyntaxError is not a ValueError/JSONDecodeError, so the except in ToolCallFormatter.__call__ did not catch it. The exception escaped into socketserver.process_request_thread, producing a noisy traceback in the server log and a half-sent response on the client side.

Fix

• Rewrite _parameter_regex with two capture groups and a trailing lookahead that anchors the closing </parameter> to one followed by the next <parameter=, the closing </function>, or end of input. Literal <parameter=...> / </parameter> substrings inside a value no longer terminate the match.
• The manual match_text.index(">") split is gone — findall now returns (name, value) directly.
• except (ValueError, json.JSONDecodeError) → except Exception in ToolCallFormatter. Future parser bugs degrade gracefully (warn + skip) instead of corrupting the response.

2. handle_completion — graceful abort on client disconnect

Problem

Long generations (large prompt cache + big model) can take many minutes. If the client times out and closes the socket mid-generation, the next self.wfile.write(...) raises BrokenPipeError (or another ConnectionError subclass). The exception bypasses the access-log layer and prints a long traceback. Worse, every retry from the same client repeated the cycle, and each retry kept the model busy generating tokens that nobody was going to read — monopolizing the generation lock and starving subsequent requests.

Fix

Add except ConnectionError (covers BrokenPipeError, ConnectionResetError, ConnectionAbortedError) right before the existing finally: ctx.stop(). The teardown then aborts the generation cleanly: ctx.stop() halts the token stream, the model becomes available, and a single concise INFO line replaces the traceback in the log.

Test plan

• python -m unittest tests.test_tool_parsing — 6/6, including the new test_qwen3_coder_value_contains_parameter_tags.
• python -m unittest tests.test_server — 20/20.
• black --check clean on all touched files.
• Manually verified against the original failing prompts: parser returns the correct value, server logs no exception when a client disconnects mid-stream, and the generation actually stops (no orphaned compute after disconnect).