Fix/43: agent tool session now cleared

[ec6862]

Summary

Fixes a session-state leak in the agent orchestrator. When the same user
requested a second review, Orchestrator.run() loaded the previous review's
stored state and merged the new results onto it with dict.update(). Any tool
that ran in an earlier review but not the current one (e.g. skill_extractor
when a résumé was removed) left stale output behind in the stored session, so
the user's re-review reflected a portfolio version that no longer existed. This
PR makes each review persist only its own results, so re-reviews start clean.

Issue

Closes #43

Changes

• agent/orchestrator.py: Removed the load-and-merge of prior session state
  (session_store.get(...) + session_state.update(results)). The current
  review's results are now written directly with
  session_store.set(profile_id, results), which overwrites the stored session
  instead of accumulating stale tools.
• tests/unit/test_orchestrator_session_reset.py: Added a regression test
  that runs two reviews for one user against a shared store and asserts the
  second review's stored session contains only its own tools.

Testing

• Unit tests pass (make test-unit)
• Integration tests pass (make test-integration)
• Linter passes (make lint)
• Type checker passes (make typecheck)
• New/updated tests cover the changes

The boxes above are left unchecked deliberately — see Notes for Reviewers. The
new regression test (test_session_state_cleared_between_reviews) fails on
main and passes with this change. My changed files (orchestrator.py and the
test) pass ruff, black, and mypy individually

Screenshots / Demo

N/A — backend logic change, covered by the unit test above.

Notes for Reviewers

• The fix is a small, contained change to how the orchestrator persists state;
  session_store.py needed no changes (its set() already overwrites the key).
• On the checkboxes: make check and make test-unit do not currently pass
  on a clean checkout of this branch's base — there are ~175 pre-existing ruff
  errors, ~5 mypy errors, and 53 failing unit tests across unrelated
  subsystems (e.g. pii_scrubber, review_service, bias_detector). I
  confirmed my change introduces zero new failures: same pre-existing set
  before and after, plus one new passing test. Happy to scope broader cleanup
  into a separate PR if desired.
• Out of scope: concurrent reviews on the same profile_id (last-write-wins).
  This PR targets the sequential re-review case described in the issue.