Hound 9.0.0

Production-hardening + repo-professionalism release. No new tools, no API breaks — the 6 tools, schemas, and response shapes are unchanged.

Fixed: clean shutdown (no more 'failed to load' on Windows)

On Windows the ProactorEventLoop's pipe transports (the warm patchright Chrome subprocess) emit Exception ignored in __del__ tracebacks to stderr during GC AFTER the event loop closes — RuntimeError: Event loop is closed and ValueError: I/O operation on closed pipe. The process exited 0, but an MCP client reading stderr saw a crash-like traceback and could report 'failed to load'. v8.2's finally block couldn't catch these (they fire in __del__ post-loop, beyond any try/except).

• serve() installs a sys.unraisablehook override that swallows ONLY the asyncio-transport-teardown noise (RuntimeError / ValueError / ResourceWarning whose traceback is in asyncio/). Every other unraisable exception still reaches the original hook, so genuine __del__ bugs stay visible.
• _shutdown_close_sessions now flushes the loop (asyncio.sleep) so pending connection_lost callbacks drain while the loop is alive, then explicitly closes any lingering asyncio subprocess transports on the loop so their __del__ is a no-op.
• Net result: python -m master_fetch exits 0 with an empty stderr on both a quick disconnect (browser mid-launch) and a long-lived session. Verified with a new CI-safe lifecycle test.

Added: CI-safe MCP lifecycle test

tests/test_v9_lifecycle.py spawns python -m master_fetch as a stdio MCP server and runs a full initialize -> tools/list -> tools/call (cache_clear) -> clean-disconnect cycle in ~3s, asserting: the handshake responds, the connect-time instructions ship, all 6 tools are present with their hand-crafted descriptions, a tool call succeeds, and the process exits 0 with no Traceback in stderr. Runs in CI (not e2e): fast and network-free.

Changed: leaner tool defs, honest token count

• mcp_smart_fetch's options description stopped advertising the 6 rare internal anti-detect knobs (real_chrome, solve_cloudflare, block_webrtc, hide_canvas, main_content_only, use_trafilatura). They still work via additionalProperties: true; the agent just isn't told to twiddle them.
• README token claim corrected from the stale '~2.5K' to the measured ~2.7K tools/list + ~0.8K one-time instructions (cl100k_base).

Fixed: stale docs + repo professionalism

• README: '9 backends' -> '10 backends' everywhere (Qwant, added in 8.1, was missing); the resilience-layer row that claimed 'Three independent indexes (DuckDuckGo, Bing, Qwant)' now reflects the real 10-backend / 6+ index-family pool; the lean-install 'keyword BM25' claim (BM25 removed in 7.2) now says 'cross-backend consensus + engine-position ranking'.
• STATUS.md (a personal agent status doc) is no longer tracked in the public repo; it is gitignored and stays local.
• Added .gitattributes (LF normalization + linguist-vendored for the ddgs-derived search_metasearch.py so it doesn't skew the language bar).
• Added community files: CODE_OF_CONDUCT.md (Contributor Covenant 2.1), SECURITY.md, PULL_REQUEST_TEMPLATE.md, an ISSUE_TEMPLATE/feature_request.yml, and ISSUE_TEMPLATE/config.yml.
• e2e test desalted: removed the dead TINYFISH_API_KEY injection (TinyFish removed in 7.0) and bumped the protocol version to 2025-03-26.

624 tests (622 + 2 lifecycle).