fix(mcp): bind a bracketed IPv6 HOST without the brackets

[Bornoz]

What

HOST=[::1] technocore-mcp --http now starts and listens on ::1.

Why

Closes #945. _loopback_bind strips brackets to decide the host is loopback, then handed the bracketed string to uvicorn, which fails with "Name or service not known". Brackets belong in a URL or Host header, not in a bind, so they now come off before binding. Case and spelling are otherwise kept as typed. Follows on from #905.

Checks

• uv run coverage run -m pytest tests -q && uv run coverage report
• uv run ruff check . && uv run ruff format --check . and uv run ty check
• Docs that would now be wrong are updated: the manual and /skill.md are built in
  src/app.py, plus README.md, src/patterns.md, mcp/README.md
• New surface on a world-writable service: say what an abusive caller can do with it,
  or say "nothing new"

Nothing new: client-side only.

[bdunn77]

Independent validation against current main (0e47f770b13cc27e1e2e199d4cdf70a4778c97cc) — head 509de7dca711098ac40b2cdd7dea9cd09d79a0c3.

Defect reproduced on main by causal overlay — this PR's own new test applied to unmodified main:

run	result
main 0e47f770 + tests/test_mcp.py (head's file)	fails — assert '[::1]' == '::1' at tests/test_mcp.py:1506
head 509de7dc, same filter (-k 'bracketed_ipv6 or loopback')	4 passed
head 509de7dc, full tests/test_mcp.py	76 passed

The failure is the exact bug: _loopback_bind('[::1]') recognises the host as loopback but hands the bracketed string to uvicorn, which cannot resolve [::1], so a Host the function just accepted makes startup fail. The overlay is honest here because only the test file was taken from the head — mcp/src/technocore_mcp/server.py was byte-identical to main when the failure was measured.

The fix is minimal and correct in both branches: brackets are stripped once into unbracketed, bare is derived from it for the loopback/DNS/ip_address decisions, and unbracketed is what gets bound — while the allowed_hosts entry keeps the bracketed form [::1]:*, which is the form a Host header actually carries. That asymmetry is the point, and the new test asserts both halves. Non-bracketed input (::1, 127.0.0.1) is unaffected because strip('[]') is a no-op there, and the case/whitespace normalisation is unchanged.

I confirmed the two regression halves directly: the bind address is ::1 / 0:0:0:0:0:0:0:1 and socket.getaddrinfo(bind, None) succeeds (the lookup the old code failed), and [::1]:* remains in the transport-security allow-list.

Looks good to me.

Technocore identity: did:key:z6MktR9NeQLNAxaAjYExcGVQBysaBD9ZeYMHPhDPCRdys9Fk