verify_uploads.py row-index reader uses splitlines() — false-FAILs the #2148 rows gate on U+2028-bearing rollout JSONL

scripts/verify_uploads.py:2463 reads fetched row-index JSONL via text.splitlines(). Python's str.splitlines() splits on U+2028 (LINE SEPARATOR), U+2029, and U+0085 (NEL) in addition to \n. Those characters occur legitimately INSIDE JSON string values in rollout text, so a valid one-record line is shredded into fragments and the fragments fail to parse.

This is the exact #1162 class the repo already banned. The existing workflow_lint.py --check-jsonl-splitlines did not catch this call shape.

Why it matters more than a cosmetic parse warning

The reader backs the #2148 realized-row-count reconciliation — the gate whose rows= token pod.py terminate REFUSES to proceed without. A false ERROR there flips the overall verdict to FAIL on healthy data, which either blocks a legitimate teardown (a wide pod keeps billing) or teaches the operator to distrust a correct gate. It fails in the expensive direction.

Observed instance

Found during issue #2546 arm-2 Step 8 upload verification (2026-08-27).

artifact   raw_completions/pre_greedy_a2/gsm8k_train.jsonl
reported   "row-index-key-absent ... Unterminated string" at line 1519
reality    file parses clean via text-mode iteration: 7375/7375 records
reality    sha256 byte-identical pod <-> HF
contains   U+2028 inside a JSON string value

The check-10 ERROR set verify_uploads.py rc=1. The verifier superseded it with an independent recount plus byte-identity, so #2546 was not blocked — but only because a human- directed recount happened to run. An unattended run would have surfaced a FAIL.

Raw evidence from that session: /tmp/verify-2546-A.json (the rc=1 run, ERROR is exactly this), /tmp/verify-2546-B2.json, /tmp/verify-2546-B3.json.

Fix

1. Replace the splitlines() read at scripts/verify_uploads.py:2463 with text.split("\n") plus an empty-line strip guard, or iterate the fetched file in text mode — whichever matches the #1162 rule's prescribed form already used elsewhere in the repo.
2. Extend workflow_lint.py --check-jsonl-splitlines so its coverage reaches this file and call shape. The lint existing but not firing here is the more general defect: any other JSONL reader with the same shape is equally exposed.
3. Add a regression fixture: a JSONL row carrying U+2028 inside a JSON string value, asserting the reader returns one record rather than two fragments.

Audit the repo for sibling splitlines()-on-JSONL readers while in there; the lint gap implies this call shape was never covered, so #2463 is unlikely to be the only one.

Provenance

Surfaced as a workflow-fix-candidate v1 by the upload-verifier subagent during #2546 arm-2 Step 8 verification, and auto-filed by the #2546 orchestrator per .claude/rules/workflow-fix-on-bug.md. Confidence: high (the mechanism is a documented property of str.splitlines(), and the false ERROR was reproduced against a file proven byte-identical and independently parseable).

Do not modify any #2546 artifact — that task's data is verified and its pod is terminated.