Free TUI test renderers that were leaking TextBuffers (CL-5539)

[TheGreatAxios]

Closes CL-5539.

Problem

CL-5539: TUI test renderers were never freed in several test files, so a full-suite run leaked native TextBuffer handles until CI ran out and started emitting "Failed to create renderer" / "Failed to create TextBuffer".

src/tui/harness.ts's withTestRenderer already frees correctly via try/finally, and most .test.ts files that call createHarness directly already pair it with a try/finally or beforeAll/afterAll destroy. Four files did not:

• src/tui/list-modal.test.ts — created a harness per test, never destroyed it.
• src/tui/provider-connect.test.ts — same, single test.
• src/tui/provider-setup.test.ts — the big one: ~30 call sites go through mountSetup/mountLogin/createHarness, but the file had exactly one harness.destroy() call. Nearly every test leaked a renderer.
• src/tui/mouse-reporting-disabled.test.ts — only freed its harness on the happy path (mountedHarnesses.pop()?.destroy() at the end of the test body), so a failing assertion above it would leak.

Fix

In each file, track every harness created and free it in a shared afterEach, so cleanup happens regardless of which assertion fails partway through a test. This matches the pattern already used by files in this suite that share one harness across a describe block via beforeAll/afterAll.

Verification

• bun test src/tui/list-modal.test.ts src/tui/provider-connect.test.ts src/tui/provider-setup.test.ts src/tui/mouse-reporting-disabled.test.ts — 67 pass, 0 fail.
• bun test src/tui (full TUI suite, 118 files / 1795 tests) — 0 fail, and grep -c "Failed to create renderer\|Failed to create TextBuffer" on the run's output is 0.
• Confirmed the leak at the code level before fixing: provider-setup.test.ts had 30 createHarness calls against 1 destroy() call; after the fix the counts balance (every harness created is tracked and destroyed in afterEach).
• bun run check (lint, typecheck, build, test) — green: 5277 pass, 0 fail across 377 files.

Left out / other issues noticed

• Did not touch product-level code — this is test-infrastructure only.
• Did not re-verify the permission-subagent-spans order dependency mentioned in the ticket's "done when" list; that needs a separate pass once this lands.
• No product bugs found during this investigation.

[linear-code]

CL-5539