vi mode: make . replay the insert command instead of injecting buffer text

[sinelaw]

Fixes #2443
Refs #2447

Motivation

Dot-repeat of the cursor-repositioning insert commands silently corrupted the buffer. o hello Esc then . on another line injected the original line's content plus a newline plus the typed text mid-word; a Y Esc then . re-injected the original line's leading character; A X Esc then . injected the entire original line's content. i and x repeated correctly, which is what localised the defect to the commands that move the cursor before inserting.

Two related defects came out of the same code path:

• O was broken before . was involved at all: the empty line opened in the right place, but the cursor landed one line above it, so the typed text corrupted the preceding line, and . afterwards reported "No change to repeat".
• An insert following an x was never recorded: x, then later i+text+Esc, then . repeated the old x.

The fix (all inside crates/fresh-editor/plugins/vi_mode.ts)

The recording model was "remember the byte span from the cursor at insert-mode entry to the cursor at Escape". That span was wrong because executeAction("move_line_end") / move_right / insert_newline are queued and applied on the editor thread, while getCursorPosition() reads the plugin's state snapshot — so the recorded start was the pre-reposition cursor, and the captured "inserted text" was really the intervening buffer text, which . then re-inserted verbatim.

Per Vim (:help .), the change is now recorded as the command plus the literally typed text, and replay re-executes the command's semantics at the new cursor:

• i/a/I/A/o/O each record which command entered insert mode (pendingInsertCommand).
• The repositioning ones await editor.flush() before switching to insert mode, so insertStartPos is the post-reposition cursor and the Escape-time capture spans exactly the typed keystrokes.
• vi_repeat's insert branch re-executes the motion (move_right / move_line_start / move_line_end / move_line_end+insert_newline / open-line-above) and then inserts the recorded text, leaving the cursor where Escape would have.
• O no longer uses insert_newline + move_up. It inserts the buffer's own line terminator at the line-start byte offset and sets the cursor there — both derived from the same pre-command cursor, so the new line and the cursor cannot disagree. The same helper is reused by ..
• Recording the entering command also fixes the x-then-i case: the capture no longer refuses to overwrite an unrelated previous change.

No Rust code changed; the plugin is embedded in the binary, so a rebuild is required to pick it up.

Tests

Added to crates/fresh-editor/tests/e2e/vi_mode_bugs.rs. They drive real keystrokes and observe only rendered output: the text rows (line-number gutter and scrollbar chrome stripped) and the status line's mode indicator and Ln N, Col N readout. Nothing reads the buffer, cursor, or editor mode out of the model. Each test gets its own TempDir project root and internal clipboard.

The assertion helper (assert_renders) waits for either the expected rows or a screen that has stopped changing, then asserts. Both arms observe state; neither is a wall-clock deadline. The second arm is what makes these useful: waiting only for the expected text means a regression never asserts at all — it runs until the external timeout kills it and reports a hang instead of naming the corruption. That was measured, not assumed: an earlier revision of the helper turned the reverted-plugin runs into stalls, and with the settle arm the same runs fail in ~1.6s quoting the exact wrong text.

With the fix

Rebased onto origin/master (clean; master had not touched plugins/vi_mode.ts, tests/e2e/vi_mode_bugs.rs, or the e2e harness). On that rebased tree, the seven tests were run by name from a freshly built e2e_tests binary: 7 passed; 0 failed.

Without the fix

plugins/vi_mode.ts reverted to master's copy, the binary rebuilt, each test run individually. Because the shared build directory is used by other work and served stale artifacts more than once, every binary used here was verified to actually contain (or, for the reverted build, to not contain) this change before being run.

Test	Keys	Result without the fix
..._dot_repeat_open_below	o hello Esc, j j, .	FAILED 1.61s — ["alpha","hello","bravo","charalpha","hellolie","delta"] vs expected […,"charlie","hello","delta"]
..._dot_repeat_append_after_cursor	a Y Esc, j 0, .	FAILED 1.66s — ["aYlpha","aYbravo"] vs expected ["aYlpha","bYravo"]
..._dot_repeat_append_line_end	A X Esc, j, .	FAILED 1.67s — ["alphaX","bravalphaXo"] vs expected ["alphaX","bravoX"]
..._open_above_cursor_placement_and_repeat	j j O hey Esc, .	FAILED 1.60s — ["alpha","bheyravo","","charlie"] vs expected ["alpha","bravo","hey","charlie"]
..._insert_after_delete_char_is_recorded	x, i Q Esc, j 0, .	FAILED 1.61s — ["Qbc","yz"] (the x replayed) vs expected ["Qbc","Qxyz"]
..._control_dot_repeat_insert_before	i AB Esc, j, .	ok 0.28s — control, pins a path that was already correct
..._control_dot_repeat_delete_char	x, ., .	ok 0.26s — control, pins a path that was already correct

Manual validation

Walked every case from the issue's reproduction comment in tmux (isolated HOME, private tmux socket) against a freshly built binary, on alpha/bravo/charlie/delta/echo. The plugin file on this branch is byte-identical to the copy that binary was built from.

Keys	Observed buffer
o hello Esc, j j, .	alpha / hello / bravo / charlie / hello / delta / echo ✅
a Y Esc, j 0 .	aYlpha / bYravo ✅
A X Esc, j .	alphaX / bravoX ✅
j j O hey Esc	alpha / bravo / hey / charlie — cursor on the opened line ✅
…then .	alpha / bravo / hey / hey / charlie ✅
x, i Q Esc, j 0 .	Qlpha / Qbravo ✅ (the insert repeats, not the x)
i AB Esc, j .	ABalpha / bABravo ✅
x, ., .	alpha → lpha → ha ✅

Other checks

• cargo fmt -p fresh-editor — no changes.
• plugins/check-types.sh vi_mode.ts — four ActionSpec.args errors, all pre-existing: the same four are reported when the checker is run against master's copy of the file. This change adds none.

Not run

Stated explicitly so the reviewer does not assume otherwise: cargo clippy and cargo check (including --features gui) were not run on this branch, and no suite broader than the seven tests named above was run after the rebase. CI covers those.

🤖 Generated with Claude Code

https://claude.ai/code/session_01D6SkGXXTcJsytTjF1TBf8Y