feat(task): use MetaMathQA (exp_metamath_p02) as the default expert group

[isabella618033]

What

Switches the default expert group for both miner and validator from exp_math to exp_metamath_p02 (dataset meta-math/MetaMathQA).

• connito/shared/config.py: TaskCfg.expert_group_name default exp_math → exp_metamath_p02.
• expert_groups/exp_metamath_p02/config.yaml: group_id -1 (sentinel) → 3 (active slot; distinct from exp_math=0 and helper exp_c4_p02=2, so no assignment clash). Group already ships an expert_assignment.json (26 layers, 146 experts).

Why it applies to miner + validator

expert_group_name is a locked field (TaskCfg._LOCKED_FIELDS), so check_and_prompt_locked enforces it against the default on every worker — changing the default propagates to all miners/validators. Run configs still pinning exp_math will be prompted/reset to exp_metamath_p02.

helper_group_id stays 2 (exp_c4_p02) — the frozen 2Fnat helper is dataset-agnostic.

⚠️ Reviewer note — sequence length / memory

exp_metamath_p02 pins sequence_length: 4096, whereas exp_math is 1024. So this switch also raises the effective seq length 1024 → 4096. Per the earlier VRAM analysis, fp32 torch.optim.AdamW OOMs at seq 4096 on a 47 GB A6000 at this model size — miners on default fp32 Adam would need either 8-bit AdamW or sequence_length: 1024. If we want the memory-safe path, drop metamath's sequence_length to 1024 in this PR — say the word and I'll add it.

Test

• TaskCfg() resolves expert_group_name=exp_metamath_p02, helper_group_id=2, locked set intact.
• exp_metamath_p02/config.yaml parses; expert_assignment.json present (26 layers).

🤖 Generated with Claude Code