Test hardening: unmocked setup/doctor e2e; drop stale setup.sh suite

[emp3thy]

Summary

Closes the two follow-ups from the self-managing-setup reviews:

• New tests/e2e/test_setup_cli.py — the first unmocked end-to-end coverage of the install path since the old install_hooks e2e was retired: runs the real better-memory setup / doctor CLI as subprocesses under an isolated home (hermetic env via isolated_env), asserting the full wiring lands (all 8 hook specs, MCP entry, CLAUDE.md managed block exactly once, sqlite default, home layout), setup is byte-idempotent, and doctor reports wiring clean (exercising the junction fallback on symlink-less hosts).
• Deleted tests/e2e/test_setup_sh.py (619 lines) — every scenario pinned removed behavior (Ollama, install_hooks flags, old script line numbers) and the module was permanently skip-marked. Canary scope repointed at the new suite; a hidden _host_uv_dirs import and two stale references cleaned up.
• Lint: wrapped the long docstring in tests/hooks/test_post_commit.py — ruff check tests now fully clean.

Test plan

• Full suite: 1817 passed, 22 skipped (8 fewer skips: the dead skip-marked module is gone).
• New e2e: 3 passed, verified against real contracts (manifest constants, setup_cmd literals, engine.diff semantics) per review.

🤖 Generated with Claude Code

[github-actions]

🟢 Claude BugBot Analysis

This PR is almost entirely test-suite churn: it deletes the legacy shell-driven tests/e2e/test_setup_sh.py, adds a new unmocked tests/e2e/test_setup_cli.py that drives the real CLI, and updates related doc comments and meta-tests to match. Verified the new test's assumptions (MANAGED_HOOKS shape/count, isolated_env signature, setup/doctor exit codes and output strings) against the real source and found no logic defects, race conditions, or security issues in the added/modified lines.

No bugs were detected in this PR.