eval: close exp_legal memorization holes (seeded pick + data-quality gates)

[george-connito]

Miners scored val_loss ≈ 0.04 vs baselines of 4.5–12.5 (perplexity ~1.04) with all miners within 0.0006 of each other. Root cause is the eval data layer, not the seed: the MinerCommit2 block-hash seed (PR #175/#177) was verified correct on chain for all completed rounds.

Validated on the live validator by reproducing the round-8670446 eval stream with production code:

• exp_legal still ran the legacy head-of-stream path (§2 shortcut of docs/exp-legal-migration-plan.md) — the whole draw came from the head of one sub-file;
• 38% of streamed Multi_Legal_Pile rows have empty text → all-padding batches → NaN loss → silently excluded from the scored divisor;
• 75% of non-empty rows share an identical 200-char prefix (templated boilerplate; zlib 0.364 vs c4 0.527);
• truncation always scored each document's first 1024 tokens — its most templated region.

Changes:

• eval_shard_pick: row_count_source="verified_table" (frozen per-shard counts double as the shard allowlist), leaf_name_pattern override, and load_builder script-bypass (Multi_Legal_Pile's builder streams files from external repos and needs trust_remote_code; the generic json builder reads the pinned native shard directly). Register (joelniklaus/Multi_Legal_Pile, all_all) @ 911e1d21 with all 29 native shards above the 10k-row headroom floor (counted by full decompress).
• dataloader: eval-only deterministic gates eval_min_text_chars=200 and eval_dedup_prefix_chars=200 (exact-prefix set — builtin hash() is per-process randomized and could break cross-validator consensus).
• tokenize_windowed: long docs contribute a content-hash-derived window instead of the boilerplate prefix (no RNG, consensus-safe). Adopted by the default dataset class and exp_legal; exp_math left for follow-up.
• exp_legal config: eval_source_seeded_shard_pick: true (shortcut retired). Miner training still streams the all_all builder mix, so eval ⊆ training distribution.
• evaluate: scored_batches/nan_batches log promoted to INFO — the only production signal that an average is silently excluding NaN batches.

Rollout: eval batches change; all validators must upgrade together (same discipline as the shuffle/skip bumps).