Description reader: rule vocabulary from outside the reader; oracle drops paths for five codes #204

Two FILE items from the external review of #183-#185 (ChatGPT batch 25, C3 and C6). Neither is a blocker.

1. Single source of rules (Spec §2.7). description_identity.py imports check_identity_text from discovery.py and SINGLE from normalization (description_identity.py:11-18).`` §2.7 says the reader takes no vocabulary from outside itself at run time. The current values are tested against the Spec (test_spec_rules_table.py:88-94), so no input is wrongly accepted or refused today; only the structure differs from the rule.
   • Fix: keep the reader's own copy, as description_rules, and test that it equals the discovery and normalization values.
2. Oracle granularity. tests/spec_oracle.py:25-52 drops the paths, and never compares values, for E_SOURCE_FOLDER_REUSED, E_SOURCE_FOLDER_CONFLICT, E_SOURCE_REPEAT_REUSED, E_MASK_SOURCE_TYPE_REUSED and E_REFERENCE_CYCLE. A refusal attached to the wrong field could therefore pass.
   • Fix: add direct reader tests that pin the offending path and value for each of these five codes.