Fix/152 faithfulness short claims not supported

[gzam1028]

Summary

FaithfulnessChecker._is_supported() required a flat >= 2 meaningful token overlap between a claim and its context, regardless of how many content words the claim actually had. Short claims (2 or fewer meaningful tokens, e.g. "Good communicator") could never be marked supported once one of those words was paraphrased in the context, since the overlap could never reach 2. This PR lowers the required overlap to 1 for claims with 2 or fewer meaningful tokens, while keeping the original floor of 2 for longer claims that still need real evidence.

Issue

Closes #152

Changes

• rag/evaluator/faithfulness_checker.py: _is_supported() now computes the claim's own meaningful token count and requires only 1 overlapping token when that count is 2 or fewer, instead of a flat 2.
• tests/unit/test_faithfulness_checker.py:
  • Added test_short_claim_with_partial_paraphrase_should_be_supported (the reproduction test for Faithfulness checker can never mark short claims as supported #152 — now passes).
  • Updated test_minimum_overlap_required's expected value ("Python expertise" vs. "Python" now correctly passes under the new 1-token floor for short claims — it previously asserted False on purpose to document the bug).
  • Added test_long_claim_with_single_overlap_stays_unsupported, test_single_meaningful_token_claim_matches, test_single_meaningful_token_claim_no_match for edge case coverage.
  • Added missing type annotations across the file and completed two pre-existing tests that asserted nothing, so pre-commit hooks pass.

Testing

• Unit tests pass (make test-unit) — see note below on pre-existing failures
• Integration tests pass (make test-integration) — not applicable, no integration surface for this change
• Linter passes (make lint) — for the files changed in this PR (ruff check on both files passes clean; the full-repo make lint fails on 180 pre-existing, unrelated errors — see note)
• Type checker passes (make typecheck) — for the files changed in this PR
• New/updated tests cover the changes

Note on pre-existing failures: This repo has known pre-existing issues unrelated to #152. I confirmed via git stash comparison that none of the following are introduced by this PR:

• test_partial_support_returns_middle_score and test_multiple_context_chunks fail due to a separate bug in _extract_claims() not splitting claims on commas/"and".
• test_none_context_chunk_text fails due to issue Faithfulness checker crashes when a context chunk has text: None #153 (crash when a context chunk's text is None).
• make lint/make check fail repo-wide due to 180 pre-existing ruff errors across unrelated files (181 without this PR's changes — this PR fixes one pre-existing import-order issue in the file it touches and introduces zero new lint errors).

Screenshots / Demo

N/A — backend logic change, no UI surface.

Notes for Reviewers

The threshold change is intentionally conservative: it only lowers the bar for short claims (≤2 meaningful tokens) and leaves the original >= 2 requirement in place for everything else, to avoid making longer/vaguer claims too easy to pass on a single lucky keyword match. See PLAN.md on this branch for the full reasoning and the risks I considered (particularly the intentional flip of test_minimum_overlap_required's expected value).