fix(agents): survive long overnight runs

- compact at memory-safe token budget (~20k) not the 131k window; gpt-oss
  hits oMLX's prefill memory guard ~47k tokens on a 24GB machine
- repair invalid_tool_calls (malformed JSON args) before echoing history —
  LangChain serializes the raw malformed string which oMLX 500s on
- retry invokes (4x) with an oMLX watchdog restart between attempts
- ALWAYS finalize: salvage picks from findings even after invoke failures

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016DvdRZkyCy9NJpHYDdPqmS

agents/compaction.py

@@ -10,8 +10,8 @@ from __future__ import annotations
 
 from langchain_core.messages import HumanMessage, SystemMessage
 
-COMPACT_RATIO = 0.9  # compact once the prompt reaches this fraction of the context window
-KEEP_LAST = 20  # verbatim recent turns to retain after compaction
+COMPACT_RATIO = 0.9  # compact once the prompt reaches this fraction of the token budget
+KEEP_LAST = 12  # verbatim recent turns to retain after compaction (keeps post-compaction small)
 _CHARS_PER_TOKEN = 4  # rough fallback when the server doesn't report token counts