Cover the personality settings write paths

[selfcontained]

Why

PersonalitySettings (399 lines) had zero coverage — no test sibling, no e2e reference, and its only caller settings-pane.tsx is untested — while owning four write paths against /api/v1/personalities. It was the top-ranked gap carried in the test-enforcer brain state.

What

One new file, apps/web/src/components/app/personality-settings.test.tsx (18 tests). Only the HTTP seam (@/lib/api) and framer-motion are mocked; every row, form, radio toggle and error banner is the real component. The api mock dispatches on ${METHOD} ${path} against a mutable list payload, so refreshes are modelled rather than stubbed away.

Pinned behavior:

• create — trimmed name, untrimmed prompt, POST then a read-after-write refresh
• edit — the form replaces that row in place, PATCH by id, no create request
• active toggle — optimistic flip before the request settles, { id: null } when the active row is toggled off, revert-to-previous on failure
• delete — window.confirm gate in both directions, DELETE then refresh
• validation — blank name and whitespace-only prompt both blocked before any write, each step flipping the banner text so a silent no-op cannot pass
• submit lock — buttons and fields disabled while a save is open, so a second click cannot double-post and mid-request edits cannot be silently discarded
• error lifecycle — surfaced on every failure path, edits preserved for retry, cleared before the post-save refresh lands, and cleared by any later successful refresh

Active state is asserted through aria-pressed/aria-label rather than Tailwind classes, so a restyle cannot break the test and a broken affordance cannot pass it.

Notes

The framer-motion mock caches its component per tag. An uncached proxy hands React a fresh component type on every render, which remounts the whole list and silently detaches any element a test is holding — that cost two confusing failures during authoring and is called out in a comment.

50 mutants were run against personality-settings.tsx. The only survivor is the personalities.length === 0 half of the loading guard: setLoading(true) happens solely in the useState initializer and refresh() only ever sets it false, so React never commits a frame with loading === true and a non-empty list. No test can kill it; it is recorded in the brain rather than worked around.

A review agent on the diff found four gaps — a non-discriminating banner assertion, an untested "edits survive a failed save" claim, and two undocumented survivors — all fixed here, with all seven demonstrated mutants re-verified as killed.

Local: check ✓ · format ✓ · lint 0 errors · unit 163/2797 server, 68/1000 web, 60 ext, 4 scripts · e2e 181 passed/12 skipped · live tmux lane 11 passed, no leaked sessions.

🤖 Generated with Claude Code