fix: make behaviour evaluation authoritative

Core fix: For cases with expectedBehaviours, reasoningQuality.status is now
set exclusively from behaviour evaluation results (required behaviour pass/fail).
Legacy concept checks remain visible as diagnostic-only metrics and do not
influence the authoritative result.

Key changes:
- Behaviour-based scoring determines reasoning status (passed/failed)
  instead of legacy concept literal matching
- Schema failure correctly forces not_evaluated (no vacuous truth)
- Saved live results re-evaluator preserves provenance metadata
- Classification tolerance map works bidirectionally for interchangeable types
- normalise() treats underscores as word characters, hyphens as spaces

Tests: 74 passing across both evaluator test suites
- tests/evaluator-behaviour-authoritative.test.mjs (47 tests, new)
- tests/evaluator-semantic.test.mjs (27 tests)

package.json

@@ -14,7 +14,8 @@
     "evaluate": "node tests/evaluator.mjs",
     "evaluate:mock": "EVAL_REAL=0 node tests/evaluator.mjs",
     "evaluate:diagnostic": "EVAL_DIAGNOSTIC=1 EVAL_REAL=0 node tests/evaluator.mjs",
-    "evaluate:live": "EVAL_REAL=1 node tests/evaluator.mjs"
+    "evaluate:live": "EVAL_REAL=1 node tests/evaluator.mjs",
+    "evaluate:saved": "node tests/evaluator.mjs"
   },
   "dependencies": {
     "next": "^14.2.0",