feat: add unit tests for model client (_parse_response, MessageBuilder, ModelConfig)

[nankingjing]

Summary

Adds 24 unit tests for the phone_agent/model/client.py module. This is the second pure-logic test suite for the repository (following PR #410 which covered parse_action).

Module tested

phone_agent/model/client.py — the AI model client responsible for parsing raw model output and building OpenAI-format message dictionaries. _parse_response and MessageBuilder are fully deterministic, data-in/data-out functions — no network or device access needed, making them ideal for fast, hermetic unit tests.

Test classes (11 classes, 24 test cases)

Class	Tests	Covers
TestParseResponseFinish	3	Rule 1: content with finish(message=...)
TestParseResponseDo	3	Rule 2: do(action=...) parsing, incl. finish-takes-precedence
TestParseResponseLegacyXml	2	Rule 3: XML <think>/<answer> fallback (strip tags)
TestParseResponseFallback	2	Rule 4: no markers / empty string fallback
TestMessageBuilderSystem	1	create_system_message shape
TestMessageBuilderUser	2	create_user_message (text-only + with base64 image)
TestMessageBuilderAssistant	1	create_assistant_message shape
TestMessageBuilderRemoveImages	3	Image stripping from messages
TestMessageBuilderScreenInfo	3	build_screen_info JSON (incl. non-ASCII preservation)
TestModelConfigDefaults	3	Default values, mutable-field isolation, overrides
TestModelResponseDefaults	1	Dataclass defaults for performance metrics

Design

• Pure and deterministic — no network, no device, no mocking required.
• Follows the exact patterns established in PR #410 (tests/test_handler.py): pytest, one test class per rule/method, docstrings.
• Documents the parsing precedence order (finish > do > XML <answer> > raw fallback).
• Documents that build_screen_info preserves non-ASCII (Chinese) app names via ensure_ascii=False.