fix: keep the cursor anchored to its text through format-on-save

[sinelaw]

Fixes #2777
Refs #2706 (the cursor half of that issue; its on-disk-state half was handled separately)

Motivation

When an on-save action rewrites the buffer (format-on-save, trim-trailing-whitespace, ensure-final-newline), replace_buffer_with_output restored the cursor as a raw byte offset clamped to the new length (old_cursor_pos.min(new_buffer_len)). Whenever the rewrite changed lengths before the cursor, that offset landed in unrelated text. Deterministic repro from #2777: a blank-line-deleting formatter with the cursor at the end of MARKER xyz (Ln 5, Col 11) dropped the cursor into omega at Ln 3, Col 3 instead of following its text to Ln 2, Col 11.

The fix

Map the cursor and any selection anchor through a diff of the pre- vs. post-format content, exactly as the issue proposes, reusing the existing line-level patience diff (fresh-core's compute_line_diff — no new dependency, no new diff machinery):

• New pure function fresh_core::diff::map_offset_through_diff(old, new, offset): walks the line hunks accumulating byte deltas; an offset outside every hunk shifts by the delta of preceding hunks (stays anchored to its text); an offset inside a rewritten region is refined by that region's char-level common prefix/suffix (so reindenting a line keeps the cursor at the end of its text), else snaps to the start of the replacement (VS Code semantics). Cost is one patience diff (which trims the common prefix/suffix first) plus O(lines) offset tables — no quadratic work on large files, and results always land on char boundaries.
• replace_buffer_with_output now computes the restored cursor and selection anchor with this mapping. The whole-buffer delete+insert mechanism and its undo behavior (Feature Request: Run Formatter on Save / Don't move view after undoing a format update #2027's leading cursor-restore event) are deliberately unchanged — only the restored positions differ. A selection whose cursor maps to the buffer end now also keeps its anchor (the restore event was previously skipped in that case).

This intentionally stays minimal per the complexity constraints: it does not convert the formatter result into minimal per-hunk edit events (which would additionally preserve markers and improve undo granularity) — that remains a possible follow-up.

Tests

All added tests fail without the fix (verified by reverting on_save_actions.rs and re-running) and pass with it:

• 8 unit tests on map_offset_through_diff in fresh-core/src/diff.rs: the Format-on-save moves the cursor to a wrong/random position (needs diff-based cursor mapping) #2777 blank-line-deletion repro, the Format On Save #2706 reindentation repro (common-suffix refinement), offsets before all edits, inside a deleted region (snap to replacement start), between two hunks (shift by the earlier delta only), no-op formatter identity, clamping, and multibyte char-boundary safety.
• 2 e2e tests in tests/e2e/on_save_actions.rs driving Ctrl+S with a grep -v '^$' formatter and asserting only on rendered output: test_format_on_save_keeps_cursor_anchored_to_its_text (status bar reads Ln 2, Col 11 after format; read Ln 3, Col 3 before the fix) and test_format_on_save_cursor_before_edit_stays_put (control: cursor before the edited region does not move).

Ran: cargo test -p fresh-core --lib (253 passed), cargo test -p fresh-editor --test e2e_tests on_save (17 passed, includes all pre-existing on-save tests), cargo test -p fresh-editor --test format_on_save_undo_view_test (16 passed — #2027 undo-view regression intact), cargo fmt, cargo clippy -p fresh-core and -p fresh-editor --all-targets (no new warnings). Also validated manually in tmux against a debug build: the exact #2777 repro now lands on Ln 2, Col 11.

🤖 Generated with Claude Code

https://claude.ai/code/session_01D6SkGXXTcJsytTjF1TBf8Y

---

Generated by Claude Code