feat: migrate LLM rule evaluation to evidence-based validation model

[Tanishq1030]

Summary

This PR separates pattern-matching evidence collection from final policy evaluation within the engine, addressing the false-positive limitations of static regex rules like ALN-001.

Instead of immediately generating a violation on an LLM API call pattern match, the match is treated as a Finding Candidate and evaluated in a post-scan validation pipeline. By default, all rules fall back to direct promotion, but specific rules (such as ALN-001/MIT-003-A) run through specialized context-aware validators to inspect the file for validation safeguards.

Closes #20

Acceptance Criteria Verification

• [GD-1] LLM API invocations are treated as candidate evidence rather than immediate violations.
  • Status: Verified. Matches are appended to a temporary list of candidates instead of generating violations instantly.
• [GD-2] Introduce a rule evaluation stage before generating findings.
  • Status: Verified. Added evaluate_candidates to act as an evaluation filter at the end of the file scan.
• [GD-3] ALN-001 findings are only emitted after evaluation determines validation is absent.
  • Status: Verified. The pipeline discards the candidate finding if it detects any validation markers (such as Pydantic, Instructor, Guardrails, or explicit # anchor: validate comments).
• [GD-4] Existing reporting format remains unchanged.
  • Status: Verified. The shape of findings returned to CLI and output templates is completely preserved.
• [GD-5] Documentation updated to describe the new evidence-based evaluation pipeline.
  • Status: Verified. Created DOCS/research/evidence_evaluation.md explaining the new pipeline design.

Proposed Changes

• Rule Evaluation Pipeline: Refactored scan_file in anchor/core/engine.py to route findings through evaluate_candidates.
• ALN-001 Context Evaluator: Added specialized logic to check for validation indicators (e.g. imports of pydantic, instructor, guardrails, marshmallow, jsonschema, .validate(), .parse_obj(), json.loads(), BaseModel, or a # anchor: validate instruction comment). Candidates with validation markers are safely discarded.
• Unit Tests: Added tests/unit/test_evaluation.py to cover both invalid and validated LLM invocation scenarios.
• Documentation: Created DOCS/research/evidence_evaluation.md explaining the new pipeline design.

Verification Results

• All 53/53 tests pass successfully (including new evaluation validation tests).

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.