Feasibility: full-graph Candle training (qlora-rs)

Decision (2026-03): keep Candle on the proxy stack (o_proj / GPT-2 c_proj + LM head) using public qlora-rs QLoraTrainer::training_step_lm over &[&QuantizedLinear] (ADR 007).

Rationale: full MHA + FFN in NF4 inside Candle would require either (a) a much larger in-tree graph aligned to every HF layout, or (b) upstream qlora-rs APIs beyond current sequential LM helper. Burn owns full-graph f32 LoRA today; Candle owns practical NF4 QLoRA on the bounded proxy.

Suffix training: CLI --qlora-ce-last-k K (default 1) applies the same embed→proxy→LM head to multiple final token positions per JSONL row, improving alignment with next-token LM on a sequence suffix without implementing full causal depth in Candle.

Revisit when: Burn ships production NF4 bases + unified adapter merge parity, or qlora-rs exposes a richer block training API without forking.