fix: mutable defaults + clean error handling in hypothesis refiner

[DavidJBianco]

Summary

Two fixes to hypothesis_assistant/hypothesis_refiner_cli.py and four other assistant entrypoints, applied in dependency order:

Commit 1 — Mutable default previous_run (supersedes #83)

Replaces previous_run: list = list() with previous_run: Optional[list] = None across all five assistant entrypoints:

• able_assistant/__init__.py
• data_assistant/__init__.py
• hypothesis_assistant/hypothesis_refiner_cli.py
• planning_assistant/__init__.py
• research_assistant/__init__.py

All call sites already guard with if previous_run: so the change is behavior-preserving.

Commit 2 — Remove double-wrap in refiner(), add clean CLI error boundary (supersedes #77)

Root cause: refiner() had an inner try/except that (a) printed to stdout (garbling parseable CLI output), (b) re-raised as a generic Exception, discarding the original type and making narrow handling impossible for any caller.

Fix:

• Remove the inner try/except from refiner() entirely. Exceptions now propagate with their native type and full traceback intact.
• Add a single try/except Exception at the CLI boundary in main() around asyncio.run(refiner(...)). On failure: prints a concise message to stderr, shows the full traceback with --verbose, exits with code 1.

This is the correct layer for this handler — a CLI's job is to not dump tracebacks at users by default, not to discriminate between autogen/provider exception types.

Closes #77 and #83 (both superseded).

Test plan

• python3 -m compileall peak_assistant — passes
• Run the refiner CLI with a bad LLM config — should print one stderr line and exit 1 (no traceback)
• Same with --verbose — should print full traceback
• Normal refinement run is unchanged

🤖 Generated with Claude Code

[DavidJBianco]

Merged into dev via fast-forward (batch-d-refiner-fixes chain).