Add EngramLite, SkipGram, BackoffNgramMixer, and Complementary Training to training stack

[Copilot]

Implements the full moonshot feature stack from PR #3's checklist. All features default OFF — no impact on existing runs.

New components

• EngramLiteEmbedding — Replaces BigramHash when ENGRAM_LITE_ENABLED=1. Multi-head hashed bigram+trigram logit bias with learned sigmoid gating. Gating is essential: raw trigrams hurt BPB (Non-record: 11L XSA-all + Full GPTQ + Selective Pruning (val_bpb=1.1154, 3-seed) openai/parameter-golf#609); gating suppresses collision noise. ~1.2M params vs ~16.7M for BigramHash at BIGRAM_HASH_SIZE=16384.

• SkipGramHashEmbedding — Additive logit bias from non-adjacent token pairs (e.g. t[-1]×t[-3]). Enabled via SKIPGRAM_HASH_SIZE>0.

• BackoffNgramMixer — Eval-time causal n-gram LM with linear-interpolation backoff and entropy-adaptive mixing. Zero artifact cost (built from already-scored tokens, never saved). Competition-compliant: full-vocabulary normalized distributions, strictly causal.

• build_bigram_stats() — Vectorized P(next|prev) precomputation from training shards (Laplace-smoothed). One-time cost at run start; not stored in artifact.

Model / training changes

• GPT._add_logit_biases() — unified helper: EngramLite XOR BigramHash, plus optional SkipGram
• GPT.token_logits() — returns (B, T, V) raw logits for BackoffNgramMixer
• GPT.complementary_loss() — down-weights tokens with high bigram probability, forcing neural capacity toward hard tokens
• SplitOptimizers — engram_lite.* and skipgram_hash.* params now routed to Muon/Adam correctly

Eval changes

• eval_val_sliding_ngram() — sliding-window eval with per-token BackoffNgramMixer mixing; activated when NGRAM_MIXER_ENABLED=1
• Final eval path checks ngram_mixer_enabled before ttt_enabled or plain sliding

Env vars

Flag	Default	Purpose
ENGRAM_LITE_ENABLED	0	Replace BigramHash with EngramLite
SKIPGRAM_HASH_SIZE	0	Enable skip-gram logit bias
COMPLEMENT_ALPHA	0.0	Complementary training strength
NGRAM_MIXER_ENABLED	0	BackoffNgramMixer at eval
NGRAM_ALPHA	0.25	Fixed mixing weight (entropy-adaptive when mode=entropy_adaptive)
NGRAM_MAX_ORDER	4	Max n-gram order

Full moonshot invocation: ENGRAM_LITE_ENABLED=1 COMPLEMENT_ALPHA=0.5 NGRAM_MIXER_ENABLED=1 NGRAM_ALPHA=0.25 NGRAM_MAX_ORDER=4

Original prompt

Pull Request: #3 continue and after verification merge and push all changes.