[Makar8000]

Fixes #16386. Sources without sentence endings (Chinese, Japanese) are sent as rows joined by lone . separator lines, and models often drop some or all of those separators and answer with one line per row. The line-count fallback added in 5cac6fb was meant to accept such a reply but never did, so every affected batch was retried and then translated row by row.

Cause

• ProcessNoSentenceEndingRow appended the next row (behind a separator) to the item it was closing, so every item but the last carried two rows. TrySplitByLineCount then dealt two reply lines to the first item, ran out of lines for the last ones, and HasSameEmptyLines rejected the reply. The same inflated item text skewed HasPlausibleProportions for these sources.
• The fallback's Chinese/Japanese branch only opened when the reply contained no . line at all, so a reply that kept some separators was rejected before any mapping was tried.

Changes

• Each item now holds its own row only; the request text still carries the separators (one line removed).
• TrySplitByLineCount drops the . lines from the reply and compares content lines with the request's, for replies that kept none or only some separators. A reply with the wrong number of content lines, merged or split rows, or a row with an internal line break answered on one line, is still rejected.
• Tests: a reply with no separators and one with a separator missing are accepted for zh and ja; merged and split replies and the line-break case are rejected; the merge step's items are pinned to their own rows. The one-separator-missing reply that MergeAndTranslateIfPossible_InvalidPeriodMarkerSplitDoesNotUpdateRows asserted as rejected is exactly the shape this issue asks to accept, so that case moved to the new accepted theory; the extra-row and doubled-separator cases stay rejected.

Verification

• LibUiLogicTests 1638 passed, 94 of them in MergeAndSplitHelperTests (8 new cases, 2 moved); SeConvTests unchanged and passing.
• Real runs of the regular llamacpp engine with Hy-MT2 7B and its official prompt on two Chinese WebVTT scripts (38 and 146 cues), requests counted through a logging proxy, same model, prompt and files before and after:

	38 cues, before	38 cues, after	146 cues, before	146 cues, after
Requests	5	4	68	9
Batch replies that dropped separators	1 of 5	1 of 3	14 of 20	6 of 9
Single-row requests	0	1	48	0
Cues left untranslated	0	0	1	0
Time	12 s	10 s	44 s	26 s

Every cue kept its timing in all four outputs. Before the fix, each batch reply with a dropped separator was rejected and its rows translated one request each; after it, those replies are mapped by line count and the row-by-row retries disappear.

• Drafted with AI assistance (Claude Code); the tests and runs above were performed as listed.

🤖 Generated with Claude Code