test: fix stale-socket e2e tests to exercise scan-based discovery

[kam-hak]

Fixes #26.

Two structural bugs in the e2e stale-socket tests prevented them from exercising what they claimed. See issue #26 for details.

Changes (tests only, no bin/ or lua/ changes):

1. test_stale_socket_recovery now genuinely exercises scan-based socket discovery. Starts nvim on --listen "/tmp/nvim.$$/0" (a path bin/nvim-socket.sh recognizes via its /tmp/nvim.*/0 glob), crashes it with kill -9 leaving a stale socket, starts a second nvim, verifies the hook rediscovers via the scan.
2. Adds opt-in scan mode to run_pretool_hook/run_posttool_hook that explicitly clears NVIM_LISTEN_ADDRESS= to force the slow path.
3. Renames test_no_nvim_graceful to test_no_matching_project_nvim_graceful and rewrites it to assert only what the body proves (clean exit + proposed file generation). Genuine isolation from ambient nvim instances would require invasive environment changes, so the test is scoped to honest claims.
4. Small corrections: assert_file_exists changed from [[ ! -f ]] to [[ ! -e ]] because unix domain sockets fail -f. setup_test_project now canonicalizes TEST_PROJECT_DIR with pwd -P to handle macOS's /tmp → /private/tmp symlink, which otherwise breaks the CWD-preference matching in find_nvim_socket.

Verified locally: 14/14 tests pass on macOS (nvim 0.11).

CI portability: /tmp/nvim.<pid>/0 is in the production scanner's glob set and works cross-platform. The pwd -P fix is macOS-specific and a no-op on Linux.

[Cannon07]

Thanks for the thorough analysis and fix!

LGTM, Merging!