faircode/SPEC.md's MCP tools table says get_benchmark_results reads paper/results-frozen/, contradicting the code and its own prose

[yakew7]

Where: faircode/SPEC.md section 11's MCP tools table, the get_benchmark_results row, vs faircode/mcp_server.py:68-72,296-298 and SPEC.md's own surrounding prose.

The gap: the table cell says get_benchmark_results wraps paper/results-frozen/results_{fairness,performance}.csv. The paragraph directly below the table, and the actual code, both say it reads the package-internal generated mirror instead:

Repro:

$ grep -n "results-frozen/results_" faircode/SPEC.md
376:| `get_benchmark_results` | `paper/results-frozen/results_{fairness,performance}.csv` | ...

$ grep -n "RESULTS_FROZEN_DIR" faircode/mcp_server.py
68:RESULTS_FROZEN_DIR = Path(__file__).resolve().parent / "_results_frozen"
70:    "fairness": RESULTS_FROZEN_DIR / "results_fairness.csv",
71:    "performance": RESULTS_FROZEN_DIR / "results_performance.csv",

RESULTS_FROZEN_DIR is faircode/_results_frozen/, not paper/results-frozen/.

Why it matters: SPEC.md is the load-bearing reference for exactly this distinction elsewhere in the same document (explaining why explainers/ and paper/results-frozen/ aren't files a real pip install faircode[mcp] ships, and why faircode/_explainers//faircode/_results_frozen/ exist as generated package-internal mirrors instead). The table cell directly contradicts that explanation, so a reader trusting the table gets the wrong mental model of where a real installed package actually reads from - the same category of gap #388 fixed for the missing mirrors themselves.

Suggested fix: change the table cell to faircode/_results_frozen/results_{fairness,performance}.csv (mirrored from paper/results-frozen/ via scripts/freeze_paper_results.mirror_for_mcp()).