feat: OAuth 2.1 for streamable-HTTP MCP servers (/mcp auth)

[Mouhand-Kaddo]

What

Interactive OAuth 2.1 login for streamable-HTTP MCP servers (auth = "oauth" in the server config), plus the test-suite root-cause fix that had been blocking this work.

Feature

• /mcp auth <name>: browser login via the SDK's OAuthClientProvider — loopback callback server on a per-endpoint deterministic port, PKCE, dynamic client registration, PRM → OASM discovery.
• Startup and reconnect are credentials-only, never blocking on a browser: valid cached tokens attach; an expired access token refreshes silently (the SDK restores stored tokens but not their absolute expiry, so FileTokenStorage stamps acquisition time and the provider seeds token_expiry_time — the SDK's own refresh path then runs). A missing or rejected refresh shows authentication required — /mcp auth <name>.
• Credentials persist per endpoint under <config_dir>/mcp-auth/ (0700 dir, 0600 files, atomic writes). /mcp logout <name> drops the session and stored credentials.
• auth = "oauth" + a static Authorization header is rejected loudly at connect (they would conflict).
• AS-controlled error text (denied consent, error descriptions) is whitespace-collapsed, control-char-stripped, and capped before reaching the terminal.
• /mcp auth prints the authorization URL in the feed so a different browser can be used; if no browser can open, the error carries the URL.
• The expected startup dead end ("No redirect handler provided") no longer splashes an SDK ERROR traceback around the TUI: automatic connects run with the mcp.client.auth loggers muted; the failure still surfaces as authentication required.

Test-suite fix — root cause of the "OAuth E2E then plain HTTP fails" flake

sse-starlette's shutdown watcher captures the uvicorn Server from the process-global SIGTERM table; when a test fixture's teardown sets should_exit and the watcher's 0.5s poll lands before that event loop closes, it flips the module-global AppStatus.should_exit (never reset by the library) — every SSE response in later tests is then cancelled at birth ("SSE stream ended without a response" / "ASGI callable returned without completing response"). Whether the poll lands in the teardown window is the flake; a loaded machine makes it look deterministic. Fixed with a per-test reset fixture in tests/conftest.py, pinned by a deterministic ordered regression pair in tests/test_mcp.py.

Tests

tests/test_mcp_auth.py (21 tests): full E2E (login, in-session refresh, restart, logout, non-interactive recovery), expired-token restart silent refresh at the attach_mcp/registry seam, rejected refresh, denied consent with control chars, no-browser fallback, loopback malformed-request and cancellation hardening; plus the interference regression pair in tests/test_mcp.py.

ruff clean. Full suite: 1105 passed — the only failures are the pre-existing tests/test_telemetry.py ones (optional opentelemetry extra not installed, files untouched by this PR).

Not yet validated against a real GlitchTip instance; the fake AS is protocol-faithful (401 → discovery → registration → PKCE → callback → refresh) but real-server quirks remain untested.