Harden operator session recovery and fallback

[Zefan-Cai]

Summary

This PR isolates TODO 3 from the larger CLI branch: harden operator session recovery, resume fallback, and attempt/session bookkeeping on top of zefan-todo-2.

What Changed

• added per-session state files under operator_state/
• added per-attempt state files under operator_state/
• record stream metadata for operator attempts
• mark broken sessions and avoid reusing them
• fall back to a fresh session when --resume fails
• added focused recovery regression tests in tests/test_operator_recovery.py

Base

• base branch: zefan-todo-2

Scope

This PR is intentionally limited to TODO 3 only.

Not included here:

• stage handoff / context compression
• paper/release package generation

Validation

• python -m py_compile main.py src/*.py tests/*.py
• python -m unittest tests.test_operator_recovery tests.test_run_manifest tests.test_stage_rollback tests.test_writing_pipeline -v