Make global search tuning reproducible

[thomasahle]

Purpose

Put the complete search-only global-tuning workflow in tools/tune/, so the running campaign can be audited and reproduced from the repository instead of private orchestration files.

This does not change sunfish.py or claim an Elo gain.

What is included

• a frozen 2:1:1 opponent panel with calibration and source/binary provenance
• weighted, reproducibly randomized panel support in logistic GP and SPSA
• the full search-knob inventory, including exact off-boundaries and conditional IID axes
• a deterministic policy gate for fuel, mate-band, and eventual-widening invariants
• immutable per-run training protocols binding engines, books, panel, space, settings, and source files
• resumable campaign orchestration and crash-safe pending-pair journals
• an honest split between auditing a frozen benchmark manifest and verifying its historical method checkout

The C twin only gains the UCI knobs needed by the shared parameter space; its defaults remain node-identical to Python.

Verification

• python3 -m unittest tools.tune.logistic_gp.test_logistic_gp tests.test_tuning_tools — 160 passed
• python3 -m unittest discover -s tools/tune -p 'test_*.py' — 10 passed
• make -C tools/ctwin test — 247 search probes and 423 move-generation lists, 0 mismatches
• python3 tools/tune/verify_recovery.py --method-root <frozen-checkout> — verified
• git diff --check — clean

[thomasahle]

Full C-twin audit found and fixed one stale test assumption.

NULL_CAP_MARGIN is intentionally independent in the lab twin, while Python currently reuses EVAL_ROUGHNESS. The old gate changed only EVAL_ROUGHNESS, so C and Python were no longer testing the same recurrence. The gate now explicitly couples the two values whenever it is checking Python parity.

Results after that correction:

• defaults: 818 probes / 901 move lists, 0 mismatches
• depth 7 subset: 235 probes, 0 mismatches
• altered QS/QS_A/roughness: 779 probes, 0 mismatches
• QS=0 focused subset: 111 probes, 0 mismatches
• altered LMR: 818 probes, 0 mismatches
• table sizes 500 and 50: 913 and 915 probes, 0 mismatches

The complete 27-position QS=0 sweep remains pending on faster hardware: it is retained in make gate, but the Python reference took over six minutes and 1 GB locally. This PR should not merge before that final sweep passes.

[thomasahle]

Added an independent prevalidation audit for the terminal consensus. It now recomputes the support/L1/SHA winner from all 20 canonical lane optima and rejects a tampered selection or evidence record. The combined tuner suite is now 161 tests, all passing.