[Infra] Tier 1 follow-up: code-reviewer concerns + honesty fixes

Context

Code-reviewer verdict on issue #36 (https://github.com/superkaiba/explore-persona-space/issues/36#issuecomment-4271309947) was CONCERNS, merge-with-follow-up. No blocking bugs landed. This issue tracks the cleanup commit.

Scope

MAJOR (documentation / honesty — not code bugs)

1. Downgrade / reword Liger log message in src/explore_persona_space/train/trainer.py and sft.py.

   • Current: module-level log tells users to install Liger-Kernel for speedup
   • Problem: Liger is disabled on every in-process path (always LoRA→PeftModel) — install instruction is misleading
   • Fix: downgrade to DEBUG OR rewrite to say "Liger-Kernel available but disabled on LoRA paths due to PEFT incompatibility (see b8dd473). Only useful for full-FT runs on distributed path."

2. Add DPO precompute memory warning in trainer.py DPO path (around :545-563).

   • When precompute_ref_log_probs=True, log a one-time INFO stating the +63% peak memory tradeoff (measured: 19.08 → 31.09 GB on Qwen-7B LoRA)
   • Helps users running memory-tight setups

3. Fix benchmark harness metric in scripts/benchmark_tier1.py and scripts/benchmark_lora_perf.py.

   • Current: reports samples_per_second for packed runs → collapses to 1/K when K examples pack into one sequence
   • Fix: add train_tokens_per_second column for packed comparisons; document that samples/sec is misleading when packing=True

MINOR (7 items from reviewer)

4. Declare packaging in pyproject.toml (used for version parsing in tokenizer shim at trainer.py:32-53)
5. Fix outdated docstring in sft.py (code-reviewer flagged a specific inaccuracy — check the review for details)
6. Remove dead helper _num_training_steps_from_trainer_state (if still unused after cleanup)
7. Fix potential IndexError in prepare_dpo_jsonl — add bounds check on empty prompt/chosen/rejected
8. Change CPU-only probe to raise TypeError not ValueError (check code-reviewer's context)
9. Make install_benchmark_callback actually idempotent (or remove the claim from docstring)
10. Document (or revert) the scope-creep refactors in commits c57ab0c and 52d6d2b that bundled _init_phase/_finalize_phase/_resolve_warmup extractions and TrainLoraConfig dataclass migration — violated "one commit per change"

Verdict post

Post <!-- epm:note v1 --> marker on issue #36 recording honestly:

• SFT ≈ 0% on LoRA (target NOT met, documented reasons: short seq, Liger-PEFT incompat)
• DPO +20% confirmed (clean win)
• Memory: DPO +63% (precompute tradeoff)
• Nothing to revert

Success criteria

• Single cleanup commit batching all 10 items (or 2-3 if grouped logically)
• ruff check . && ruff format . clean
• Existing tests pass
• Commit message: chore(tier1-cleanup): address code-reviewer concerns [refs #36]
• Posts <!-- epm:cleanup-done v1 --> marker on THIS issue

Budget

≤1 hr, no GPU needed.