test(langchain): add unit tests for AgentMiddleware — core runtime has no test coverage

[caoergou]

Problem

AgentMiddleware is the core runtime of UniHarness (libs/uniharness/uniharness/langchain/middleware.py). It coordinates the full pre-model pipeline including:

• Compaction 3-phase state machine (NONE → REQUESTING → APPLYING → NONE)
• Permission gating (awrap_tool_call with PermissionGate)
• Skill injection (detects Skill tool calls, appends skill content)
• Image extraction for non-Anthropic providers (_extract_tool_images)
• System reminder rules (annotates the last message before model)

However, the tests/unit_tests/middleware/ directory only contains an empty __init__.py with no tests whatsoever:

```
tests/unit_tests/middleware/
└── init.py # empty
```

For comparison, other modules all have accompanying tests:

• harness/ → tests/unit_tests/harness/
• tools/ → tests/unit_tests/tools/
• mcp/ → tests/unit_tests/mcp/

This is a significant gap given that AgentMiddleware is the most critical runtime component.

Suggested Test Coverage

The following behaviors should be covered:

1. abefore_agent system prompt injection (idempotency) — injects SystemMessage on first turn; does not duplicate if already present
2. abefore_model compaction REQUESTING phase — appends compaction prompt without touching existing messages
3. _extract_tool_images — extracts image blocks from ToolMessage into follow-up HumanMessage; returns None when no images; skips already-processed messages (idempotency via _IMAGE_EXTRACTED marker)
4. _detect_skill_call — correctly identifies the most recent Skill tool call; returns None when a user message already follows (already injected)
5. _supports_tool_images — returns True for ChatAnthropic, False for other models
6. awrap_tool_call DENIED — returns error ToolMessage, does not call the handler
7. awrap_tool_call ALLOWED — calls through to the handler and returns its result
8. aafter_model compaction trigger — triggers when token count exceeds threshold

Environment

• libs/uniharness/uniharness/langchain/middleware.py — the module under test
• Test infra: pytest-asyncio with asyncio_mode = "auto", mypy --strict, ruff
• Pattern: test_<action>_<condition>_<expected_result>

I'd be happy to submit a PR adding these tests if this direction looks good.