Math-validation MED fixes (docs/research/2026-06-11-math-validation.md):
- Sign election is now always the magnitude-weighted 'total' vote —
the only mechanism the TIES paper (NeurIPS 2023) defines
(gamma = sgn(sum of trimmed task vectors)). The frequency vote was a
non-paper extension that served as the default; it remains available
only as an explicit merge-strategy override. The empirically-tuned
density clamp is deliberately left untouched.
- v2 composite normalizes effective rank by the stack's achievable rank
budget (sum of per-LoRA average ranks) instead of a fixed /40: a merge
of two rank-16 LoRAs can never exceed effective rank 32 (Roy-Vetterli
erank <= true rank <= summed source ranks), so the fixed cap scored
different stacks on different scales. v1 keeps /40 (frozen legacy).
Bump AUTOTUNER_ALGO_VERSION 1.8.0 -> 1.9.0 (ranking-relevant).
Adds sign-election test (245 pass).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
