Fix 18 latent splitlines-shreds-JSONL sites in experiment code (from #1162 live-tree enumeration)

Overview / Motivation

Filed from task #1162 (plan §4.6/§4.8 + the fact-checker's independent re-derivation, 2026-07-09). While extending the check_jsonl_splitlines lint, a live-tree enumeration found 18 latent splitlines-shreds-JSONL sites in EXPERIMENT code — all outside #1162's workflow-surface scope, all evading the lint's six signals by design (path-variable / cross-function dataflow in non-globbing modules; the lint's documented deliberate false negatives).

Each site reads a .jsonl file via .splitlines(), which splits on raw U+2028/U+2029/NEL inside ensure_ascii=False JSON strings — silent record drop on tolerant readers, JSONDecodeError on strict ones, inflated row counts on len() asserts (#825/#950; .claude/rules/gotchas.md). Spot-checked sites are genuine .jsonl reads (true positives), e.g. issue825_onpolicy_u2_gen.py:531 row-counts a file its own writer emits with ensure_ascii=False — the exact #825 inflated-count class.

Priority note: 3 of the 7 files sit in the ACTIVE #612/#906/#1090 datagen reuse path (behavior_testbed_545/corpora.py, elicit_v2.py, sycophancy_onpolicy_612/build_onpolicy_pool.py) — a shredded on-policy pool row is silently dropped or crashes a strict reader.

Goal

Fix each site to split("\n") + if line.strip() guard (or text-mode file iteration), per the #950 recipe — one mechanical fix class across 7 files.

Sites (fix each; line numbers as of 2026-07-09 — re-locate by pattern if drifted)

Same-scope path-variable assignment shape (16):

1. scripts/issue545_train_cell.py :188, :189
2. scripts/issue825_onpolicy_u2_gen.py :531
3. scripts/issue_653/i653_dispatch.py :1215
4. scripts/run_issue475_cot_install.py :124, :130
5. src/explore_persona_space/experiments/behavior_testbed_545/corpora.py :603, :673, :744, :883, :1561, :1621, :1676
6. src/explore_persona_space/experiments/behavior_testbed_545/elicit_v2.py :833
7. src/explore_persona_space/experiments/sycophancy_onpolicy_612/build_onpolicy_pool.py :723, :734

Cross-function parameter-receiver shape (2): 8. scripts/issue545_train_cell.py :342 9. src/explore_persona_space/experiments/sycophancy_onpolicy_612/build_onpolicy_pool.py :577

Constraints / notes

• Experiment code — this is NOT a workflow-fix task (out of .claude/rules/workflow-fix-on-bug.md scope by definition); ordinary code-fix pipeline applies.
• Frozen per-issue scripts (items 1-4) are candidates for JSONL_SPLITLINES_LEGACY_ALLOWLIST treatment instead of fixes if their tasks are terminal — but the 3 actively-reused datagen files (items 5-7) should be FIXED, not allowlisted.
• Verify with a targeted grep + the existing lint (the fixed sites should simply stop matching .splitlines()); no behavioral re-runs needed (fix is read-path-only, output-identical on ASCII-clean data, strictly more correct on U+2028-bearing data).

Provenance

• Enumerated by #1162 plan §4.6 (planner scan + independent fact-check re-derivation, both 2026-07-09); carried in #1162's epm:results v1 marker per plan §4.8.