test(tui): follow the breadcrumb and the block splash mark

[sosidudku1]

Seven tests were failing on main against code that behaves correctly. They assert on a UI surface that two merged PRs replaced.

What changed under them

• feat(tui): ctrl+p operator menu, and ctrl+g chords to jump straight there #171 replaced the three-section pill row (▸ Observe) with a breadcrumb, so the screen now reads Observe ▸ Feed. The section label is the left half of a crumb rather than a standalone pill.
• feat(tui): start page adapts to terminal size #151 made the splash mark a scaled block raster (█), so the ::: glyphs the smoke test looked for no longer appear at any size.

Why this is a test fix and not a code fix

The behavioural assertion in the same file — does not quit when Esc is pressed twice from an Observe tab, which counts quit/abort callbacks rather than matching text — passed throughout. Esc has been returning to Run correctly the whole time; only the strings moved.

The changes

• escape-observe-tabs.test.tsx — ▸ Observe → Observe ▸, ▸ Run → Run
• escape-import-tab.test.tsx — ▸ Manage → Manage ▸, ▸ Run → Run
• tui-app.test.tsx — the splash assertion now checks the rail the Run screen always carries, since the mark's glyphs vary with the surface

Tests only; no production code is touched.

Delta

v0.2.2 baseline: 7 failures / 4075 tests. Before this PR: 14 / 4309. After: 5 / 4309.

The five that remain were all failing on v0.2.2 too and are out of scope here: two TuiApp (smoke) cases (Shift+Tab expects Telegram as the last Manage tab, now Privacy; two-mode LLM panel expects a heading the panel no longer renders), plus persistEmbeddingHybridRecall, the Sibiliainen CV search case, and the sidecar FIFO concurrency test. Esc from an idle Manage panel also appears intermittently — it passes in isolation and settles ~10ms after the write, under Ink's 20ms flush delay.

🤖 Generated with Claude Code

[sosidudku1]

Pushed two more fixes so this closes out the stale-expectation cleanup completely rather than leaving two behind.

• Shift+Tab from Run expected Telegram as the last Manage sub-tab; MANAGE_TABS gained import and privacy since, so the wrap target is Privacy. The assertion carries the ▸ active marker — a bare "Privacy" also matches the inactive chip in the strip, which made the first version of this fix pass even with privacy removed from the list.
• The two-mode LLM panel asserted RouteCard chrome that correct code deliberately drops: ink-testing-library reports no rows, so LlmPanel falls back to 80x24 and picks its compact header. The full/compact decision is already owned by components/llm-panel.test.tsx, which drives maxRows explicitly, so the smoke test now asserts the two-mode body every budget keeps.

Both were checked by ablation — each fails when the behaviour it describes is removed.

Suite on this branch over current main (which now includes #165): 4766 tests, 2 failures, down from 5. The two left are the flaky sidecar FIFO case and the $HOME-dependent glob test, both failing on pristine main and out of scope here.