fix(coding-agent): preserve line endings when apply_patch updates a file

[code-yeongyu]

Fixes #2638. The standalone package's same fix is code-yeongyu/pi-apply-patch#48; senpi's builtin is a hand-maintained port (MANUAL_PACKAGES), so it gets its own fix.

What

• The bug: updates split files on LF and joined them back with LF, so a one-line patch rewrote every line of a CRLF or mixed-ending file. That showed in git, and in the tool's preview as a whole-file diff.
• line-endings.ts (new): SourceText keeps each line's own ending (\r\n, \n, lone \r). replacementsAroundContext splits a matched chunk around its context lines, so those stay exactly as in the file. Inserted lines take the file's first ending, and every line still ends with one.
• Where the codebase changes: replaceChunks uses it, and the parsers record each chunk's contextLineIndices.
• Codex comparison: this is Codex's PreserveLineEndings model (text_file.rs, compute_replacements). In Codex it is behind an under-development flag, off by default, and its conformance scenarios run with it on. oh-my-pi restores the detected ending after edits.
• Two user-visible changes: endings are preserved, and a context line matched only after trimming is left as it was instead of being rewritten to the patch's copy. LF files produce identical bytes.

Tests (test/suite/gpt-apply-patch-line-endings.test.ts)

TEST AUTHORING GATE:

1. What it protects: the update file-content contract (only changed lines are rewritten).
2. The regression it catches: a return to LF-normalizing whole-file rewrites.
3. Why existing tests miss it: no owner test uses CRLF or mixed-ending sources.
4. New production seam: none; the tests use applyPatchDetailed and the tool's execute.

Case	main	This PR
CRLF change plus insertion	fails	passes
Mixed endings	fails	passes
Trimmed-match context line untouched	fails	passes
One-line CRLF change previews as +1 -1	fails	passes
LF file without a final newline	passes	passes (regression guard)
Failed hunk leaves CRLF bytes untouched	passes	passes (regression guard)

Verification

• New file 6/6. test/suite/gpt-apply-patch* plus test/permission: 25 files, 565/565.
• Repo tsc --noEmit clean, biome clean, changelog gate PASS.
• Codex's 26 portable scenarios replayed against the builtin: 023 and 024 now pass. 24/26; 017 is fixed by fix(coding-agent): reject stray lines between apply_patch file sections #2637.

---

Summary by cubic

Fixes apply_patch rewriting every line of CRLF and mixed-ending files as LF on update, so a one-line patch no longer becomes a whole-file diff in git and in the preview (#2638).

Files are now parsed into lines that keep their own ending; only touched lines are replaced, context lines stay exactly as they were, and inserted lines take the file's first ending. A context line matched only after trimming is also left untouched. LF files produce identical bytes.

Pins pi-apply-patch to 0.1.4, which carries this fix and the indented-header fix from #2637 upstream; the grammar declaration it added is already native here.

Tests

• Eight new cases cover CRLF, mixed-ending, and trimmed-match context handling, LF and failed-hunk regressions, and CRLF files combined with the indented-header and stray-line rules of apply_patch silently drops a file section whose header is indented, and reports success #2636.

^{Written for commit 8ac5ce9. Summary will update on new commits.}

[code-yeongyu]

Lead in-session review at 1814736: PASS. Same model as pi-apply-patch #48. SourceText.parse splits on exactly the endings normalizePatchText used to fold (CRLF, lone CR, LF), so dropping that call from replaceChunks changes only the output endings, never matching. Both the batch parser and the streaming parser record contextLineIndices on the same two branches (' ' and the bare empty line), so context lines keep their source text and ending on both paths; inserted lines take the file's first ending; LF-only output is unchanged. Lands after the 2026.10.4 dispatch with #2637.

[code-yeongyu]

Real-CLI QA (senpi-qa Channel 3 pattern, zero tokens). The real senpi --print CLI runs in an isolated sandbox against the local fake model server, with a scripted gpt-5.5 model whose only tool call is apply_patch. The real credential file's hash was unchanged.

Scenario	main (control)	this branch (1814736)
CRLF lines.txt (one\r\ntwo\r\nthree\r\n): change one line, insert one	FAIL: rewritten to ONE\ntwo\nbetween\nthree\n	PASS: ONE\r\ntwo\r\nbetween\r\nthree\r\n