fix: rewrite use_rslora literal assert in test_issue1090_fu5_round (red on main)

Overview / Motivation

Filed by the #1660 orchestrator from an implementer-surfaced finding (2026-07-24): a test red on pristine origin/main, polluting every Step 9c / merge-gate baseline until fixed. NOT a workflow-surface fix — experiment-pinning test code.

Goal

Make tests/test_issue1090_fu5_round.py::test_fmt_pers_r256_train_config_carries_rank assert the rsLoRA contract in a literal-drift-proof way so it passes on main while still pinning the behavior it was written to pin.

Bug

• Observed: the test asserts "use_rslora=True" in inspect.getsource(train_lora), but main's train_lora now threads use_rslora=cfg.use_rslora (sft.py L1489) with the default carried by the TrainLoraConfig.use_rslora: bool = True field (L605) — the literal survives only in a comment (L601), OUTSIDE train_lora's source. The assert fails on pristine main.
• Why it matters: a permanently-red main test degrades the known-red ledger and burns per-branch triage; it also contradicts the sibling pin tests/test_train_lora_config_use_rslora.py, which already pins the field-based contract.
• Fix shape: rewrite the assert to the field-based contract — TrainLoraConfig().use_rslora is True plus the threading pin "use_rslora=cfg.use_rslora" in inspect.getsource(train_lora) — preserving the test's r=256/alpha=64 rank assertions unchanged.
• verified-at-filing: grep -n "use_rslora" src/explore_persona_space/train/sft.py → literal use_rslora=True only at L601 (comment); live probe uv run python -c "import inspect; from explore_persona_space.train.sft import train_lora; print('use_rslora=True' in inspect.getsource(train_lora))" → False (2026-07-24). Failing assert at tests/test_issue1090_fu5_round.py:204.

Scope

• tests/test_issue1090_fu5_round.py only (the assert at L204 and any sibling literal-asserts in the same test). Do NOT touch src/explore_persona_space/train/sft.py — the behavior is intact and separately pinned.