fix(llm): make ChatContext.truncate() honor a non-positive max_items - #7367

What

ChatContext.truncate(max_items=...) used self._items[-max_items:]. In Python -0 is 0, and items[0:] is the whole list -- so truncating to a budget of zero items returned every item, and a negative budget silently trimmed from the front of the conversation instead of keeping nothing.

The guard now treats any non-positive budget as "no conversational items", which still leaves the preserved instruction message, exactly as the docstring describes.

before:  truncate(max_items=0)  -> ['system', 'user', 'assistant', 'user']   (nothing dropped)
after:   truncate(max_items=0)  -> ['system']
before:  truncate(max_items=-2) -> front-trimmed slice
after:   truncate(max_items=-2) -> ['developer']

Why it matters

truncate is the documented way to bound a context (llm/chat_context.py:571). A budget normally comes from a token/message calculation, and the interesting case is the one where it reaches 0: the caller asks for no history and gets all of it, so the request goes out at full size instead of failing or shrinking -- the opposite of what the bound was for. #4843 / #4845 earlier fixed which roles count as an instruction and the duplicate insert, but the slice itself was untouched, so this path is still live.

Valid (positive) budgets are unaffected: len(self._items) <= max_items still short-circuits, and items[-max_items:] is unchanged for max_items > 0.

Tests

Three regression cases in tests/test_chat_ctx.py, next to the existing truncate tests. Differential, with only chat_context.py reverted:

FAILED tests/test_chat_ctx.py::test_truncate_zero_max_items_keeps_only_instruction
FAILED tests/test_chat_ctx.py::test_truncate_zero_max_items_without_instruction
FAILED tests/test_chat_ctx.py::test_truncate_negative_max_items_keeps_only_instruction
3 failed, 6 passed, 63 deselected

with the fix: 71 passed, 1 skipped   (whole tests/test_chat_ctx.py)

ruff check and ruff format --check are clean on both files.

AI disclosure

This fix was produced with an AI coding assistant (Claude), reviewed and verified by me locally as described above.