fix(agent): reset session state between reviews

[HissanT]

Summary

This PR fixes the stale agent session-state bug where a later review for the same profile could keep tool results from an earlier review. The orchestrator now starts each review run with a fresh in-memory context cache and saves only the current review's tool results to the session store.

Issue

Closes #43

Changes

• Added focused unit coverage for review session boundaries in tests/unit/test_orchestrator_session_state.py.
• Reset the orchestrator's in-memory ContextManager at the start of each review run.
• Stopped loading and merging previous persisted session state into the current review.
• Saved only the current review's tool results back to the session store.
• Updated docs/JOURNAL.md with Week 9 check-ins and validation notes.

Testing

• Unit tests pass (make test-unit) with documented pre-existing failures/no new failures
• Integration tests pass (make test-integration)
• Linter passes (make lint) on touched files; full repo has documented pre-existing failures
• Type checker passes (make typecheck) with documented pre-existing failures/no new touched-file failures
• New/updated tests cover the changes

Focused validation:

• .venv\Scripts\pytest.exe tests\unit\test_orchestrator_session_state.py -v -m unit -> 2 passed
• .venv\Scripts\ruff.exe check agent\orchestrator.py tests\unit\test_orchestrator_session_state.py -> passed
• .venv\Scripts\black.exe --check agent\orchestrator.py tests\unit\test_orchestrator_session_state.py -> passed

Full-suite note:

• make is not available in my local PowerShell environment, so I ran the matching virtualenv commands directly.
• Before this fix, the full unit suite had 53 failures.
• After this fix, the full unit suite still has 53 failures, while the 2 new orchestrator tests pass.
• Full ruff, black, and mypy checks also had pre-existing failures across unrelated files.

Screenshots / Demo

Not applicable.

Notes for Reviewers

Please focus on agent/orchestrator.py and tests/unit/test_orchestrator_session_state.py. The intended behavior is that each new review gets fresh tool state, while memoization remains available inside the current run only.