test(e2e): drive current modal surfaces and Escape-based cleanup, fixing multi-value and PROMPT-02

[chhoumann]

Repairs the two failure classes in #184 - and corrects its diagnosis: the prompt/suggester UI renders fine in the headless isolated instance. The failing tests were waiting for modals the product no longer opens.

Root cause. Since #168, editMetaElement routes eligible top-level YAML properties (Auto Properties are intercepted first) to the NativePropertyPrompt (Obsidian's native widgets). The 7 multi-value tests and audit-modals PROMPT-02 were still driving the pre-#168 GenericSuggester + GenericPrompt flow, so their .suggestion-item / .metaEditPromptInput waits timed out with the native modal open - and each stuck modal cascaded modal-count-leak failures into later files. (That the legacy modals render headlessly was already proven green by audit-suggester.test.ts, which clicks real .suggestion-item rows and types into .metaEditPromptInput, and by metaedit-runtime.test.ts's nested-YAML prompt flow; the new inline-Dataview tests below pin it in this suite too.)

What changed:

• multi-value rewritten for the current surfaces. The Metaedit doesn't edit arrays of tags properly. #94/works great apart from the tags YAML front matter, can we get a list of tags and add multiple tags? #51/comma/cmd: invariants now drive the native prompt's real widget interactions - in-place pill edit (focus pill, Enter, retype, Enter; order preserved), pill add, contenteditable text - with the same exact-YAML assertions as before. The scalar tests additionally pin that neither legacy modal participates. The legacy GenericSuggester+GenericPrompt editor still owns inline Dataview fields, so four new tests pin that surviving surface, including the add-command discrimination (VALUE_SELECTION_PREFIX vs sentinels) that only runs there.
• PROMPT-02 asserts the native date widget (input[type=date].mod-date inside the native host) instead of the legacy prompt's date input, keeping its original assertions on the surface a user actually gets.
• Escape-based modal cleanup, centralized and failure-safe. Obsidian 1.13 removed .modal-close-button, silently turning every class-targeted close click into a no-op (same family as test(e2e): survive Obsidian 1.13 settings popout and modal DOM changes quickadd#1523). A shared closeAllModals (harness-exported, in-app) closes via Escape through the keymap scope - running real onClose paths so promises resolve and scopes pop - with app.workspace.activeModal?.close() as lifecycle-aware fallback and a loud diagnostic error instead of DOM removal. It replaces every .modal-close-button click and raw el.remove() teardown across the audit/runtime/auto-properties/bulk/fluid suites, and is wired into the harness's beforeDataRestore, so a FAILED test's leaked modal is unwound before the next test - killing the cascade class at the root.
• Shared native-prompt drivers. native-properties' in-app helpers moved to tests/e2e/nativePromptHelpers.ts (now also used by multi-value), gaining editPillByText and a suggest-popover guard: typing in a multi-select opens a fuzzy popover whose scope owns Enter, which would commit the highlighted suggestion instead of the literal text (this was live: adding tag "asia" committed "course-flash/writing-obsidian-plugins").
• Vault-agnostic typing. List/scalar drives pin the key's property type via metadataTypeManager.setType (as native-properties already does), because widget resolution follows vault-wide type memory - in the shared dev vault, authors is typed text by unrelated notes and would mount the text widget.

Verification (Obsidian 1.13.0, isolated instance per AGENTS runbook):

• Full e2e suite on a fresh instance: 16 files, 106 passed, 1 skipped (documented manual), 0 failed - vs the issue's 94 passed / 8 failed baseline.
• Failure-safe cleanup proven directly: a temporary test that opened a native prompt and threw left the next test observing modalCount === 0 (temp test removed after verification).
• pnpm run lint, pnpm run build, pnpm run test (355 unit tests) all green.

No product code touched; no release impact.

Fixes #184