fix(apply_diff): detect malformed ------- separator and surface helpful error (#12210)

[SAY-5]

Closes #12210.

Summary

Smaller models (Qwen 3.6 in the issue) sometimes emit an apply_diff payload where the ------- separator is missing its trailing newline:

<<<<<<< SEARCH
:start_line:7
-------import { useTranslate } from '...';
...

The outer regex makes the separator group optional, so the run-on line is silently absorbed into the search content. Roo then reports:

No sufficiently similar match found at line: 7 (63% similar, needs 100%)

which sends the model into a loop, trying the same malformation again. The reporter notes the model can recover if told what's wrong — so a clearer error fixes the cause of the loop.

Change

Add detectMalformedSeparator(searchContent) that runs against each replacement's search content right after marker unescaping. Flags first lines that start with seven-or-more dashes followed by non-dash, non-whitespace content on the same line. Returns the offending line so the error message can quote it; otherwise returns null and the existing flow continues unchanged.

When malformed, emit:

Malformed separator in SEARCH section

Debug Info:
- The "-------" separator that follows :start_line:/:end_line: must be on its own line, followed by a newline before the search content begins.
- Got: "-------import { useTranslate } from '...';"
- Expected:
    :start_line:7
    -------
    <search content here>
- Tip: insert a newline immediately after "-------". Do not put the first line of search content on the same line as the separator.

False-positive guards: a search line that legitimately contains a long row of dashes inside the file content (markdown HR, comment header like /* ---------- header ---------- */) is NOT flagged because the helper requires a non-dash, non-whitespace tail.

Test plan

Three regression tests under malformed \-------` separator detection`:

• Exact repro from the issue → "Malformed separator" + "must be on its own line" surfaced; the unhelpful "63% similar" message no longer appears.
• Well-formed -------\n separator still applies cleanly.
• SEARCH content with a long row of dashes inside the file is NOT misclassified.

Results:

• pnpm vitest run core/diff/strategies/__tests__/ → 66 passed; 0 failed (3 new + 63 existing).
• pnpm run check-types → all 14 tasks pass.

Notes

• No production-path behavior change for well-formed diffs.
• Per-replacement (inside the for (const replacement of replacements) loop), so other replacements in the same diff still apply normally if only one is malformed.
• Changeset added under .changeset/fix-apply-diff-malformed-separator.md (patch).

Interactively review PR in Roo Code Cloud