Fix int() crash on float strings and model alias suffix stripping

[sridhar-3009]

Fixes #302, #308, #312

Changes

Fix 1: _record_tool_carryover() crashes on float-format numeric strings (fixes #302)

File: src/openharness/engine/query.py

Some LLMs emit JSON integer fields as floats (e.g. 250.0 instead of 250). int() raises ValueError on such strings. Wrapping with float() first handles both integer and float-formatted numeric strings:

# Before
offset = int(tool_input.get("offset") or 0)
limit = int(tool_input.get("limit") or 200)

# After
offset = int(float(tool_input.get("offset") or 0))
limit = int(float(tool_input.get("limit") or 200))

Fix 2: Async subprocess stdout data loss between chunks (fixes #308)

File: src/openharness/tasks/manager.py

_copy_output() writes 4 KB chunks to disk but never calls flush(). If the subprocess exits abruptly between iterations, data sitting in the Python file buffer is lost. Added handle.flush() after handle.write(chunk).

Fix 3: MiniMax [1M] model alias suffix rejected by API (fixes #312)

File: src/openharness/config/settings.py

resolve_model_setting() was passing display aliases like MiniMax-M3[1M] raw to the provider API, causing a 400 unknown model error. Added a regex strip for any trailing [N K/M] context-window suffix before returning the resolved model ID. This is consistent with how Claude handles similar display variants.