[gnapse]

A review of the existing tests found eight that would still pass with their feature broken. This PR makes each one fail when its feature breaks. Only tests changed. No real bug turned up.

What changed

• src/editor/layers.test.ts
  • The jump test writes its offsets (6, 35) out and checks the page text they refer to. Before, it took them from versionedItems, which jumpToItem itself calls.
  • The settleAll test checks the layers. Before, it checked findAlt(state, "x"), an id that never exists.
  • The in-place edit test checks the text and saved Markdown after each cycle, not only at the end.
  • The fork and drop-version tests check the saved Markdown after undo. The fork test also checks it after redo.
• src-tauri/src/lib.rs: the save test reads the file back. It expects the literal "changed_on_disk" that src/lib/files.ts matches, not the code's own constant.
• src-tauri/src/ai.rs: the request body test compares the whole body: model, system, prompt, max_tokens, effort, and schema.
• src-tauri/src/claude_code.rs: the arguments test, renamed to builds_the_command_line, compares the whole list.

Evidence

Each bug was planted in the code. The old test (from main) and the new one ran against it, then the code was restored.

#	Planted bug	Old test	New test
1	versionedItems puts alternatives at their end (from: a.to)	passes	fails: [9, 0], expected [6, 0]
2	settleAll leaves one-variant alternatives	passes	fails: layersOf still has alternatives
3	cycleAlt does nothing ((alt.active + n) % n)	passes	fails: the strugglse here, expected the tension here
4a	undo resets a version's active to 0	passes	fails: saved Markdown stars the wrong version
4b	undo keeps only the shown version	passes	fails: saved Above.\n\nA.\n
4c	redo doesn't restore the layers	passes	fails at the redo check (layers.test.ts:169)
5	write_text returns Ok without writing when the file is unchanged	passes	fails: file holds "mine", expected "edited"
6	CHANGED_ON_DISK renamed to "changed-on-disk"	passes	fails: Err("changed-on-disk")
7a–c	max_tokens fixed at 4096 / prompt sent as system / schema dropped	passes	fails: the body differs in that field
8a–c	--system-prompt / --safe-mode / --strict-mcp-config removed	passes	fails: the list lacks that flag

Two fresh reviewers ran with the review-changes skill: a task review and the .agents/checks/tests.md check. Both passed the change, and between them they reproduced rows 1–6, 7b, and 8b. pnpm verify passes.

Choices to look at

• ai.rs writes the model id claude-opus-5-5 out. Changing MODEL now means changing this test too. That is on purpose: the API relies on the value.
• claude_code.rs checks the exact flag list, in order. It catches a flag that was added (say, one that turns tools back on), not only one that was removed. The cost: reordering the flags breaks the test.

Follow-ups (not in this PR)

• No test undoes and redoes forkVersion, settleAll, removeAltVariant, or deleteGhosts, though CLAUDE.md asks for one for each command that changes layers. A reviewer planted addToHistory.of(false) in forkVersion and in settleAll, and all tests still passed.
• No test cycles backward (cycleAlt or cycleAtCaret with -1), which ⌥K uses. A cycleAlt that always steps forward passes every test. A backward test needs three or more variants.