fix(config): reject non-list groups instead of coercing

[blackaxgit]

Bug

groups = "web" in servers.toml (a string, not a list) loaded as ('w', 'e', 'b'). ServerRegistry._load called tuple() on the raw value before Pydantic saw it, so:

• the config started with three "undefined group" warnings;
• the server silently dropped out of group web;
• execute_on_group skipped it.

Other non-list values were also coerced:

• groups = 5 escaped as a raw TypeError;
• groups = { web = 1 } loaded as ('web',).

Reproduced on main before the fix.

Fix (minimal)

• config.py: delete the pre-conversion. Pydantic already turns a TOML list into tuple[str, ...].
• models.py: add a mode="before" validator on ServerConfig.groups that rejects any non-list. Pydantic's own message is "Input should be a valid tuple", which a TOML author can't act on, so the refusal now reads 'groups': Value error, must be a list of group names, e.g. groups = ["web"]. Validation stays in the model, as AGENTS.md requires.
• tests/test_config.py: one regression test, parametrised over "web", 5 and { web = 1 }. All three cases fail on main: two do not raise, one raises a TypeError.
• CHANGELOG.md: a Fixed entry, plus a note that the next release is 0.9.0: two configs that used to start now refuse to start, which the rule recorded under 0.8.1 treats as a minor bump.

A separate commit fixes doc drift found while validating #66, checked against the installed mcp 2.2.0:

• AGENTS.md: the _build_http_app signature quoted there still had a host parameter that no longer exists;
• server.py: two comments and one error string still said mcp 2.1.1 or self.session_manager.

Verification

• HYPOTHESIS_PROFILE=ci pytest: 853 passed. ruff format and check, mypy, bandit, pip-audit and gitleaks are all clean.
• Smoke through ServerRegistry and ssh-mcp healthcheck:
  • ["web"] loads as ('web',) and is a member of web;
  • [] loads as ();
  • "web", 5 and { web = 1 } each give a ConfigError that names the fix;
  • [1] still reports 'groups.0'.
• Review skill, two rounds:
  • Round 1: fable 5.1 (Claude Code code-review) raised 2 should-fix and 4 nits. sol 6.1 (codex review --uncommitted) raised 1 nit.
  • Round 2: fable raised 1 should-fix (the message) and 3 nits. sol had no findings.
  • Everything was applied except two suggestions: **server_data (the reviewers disagreed) and a name-key issue that predates this change.
• Three-seat panel (fable 5.1, sol 6.1, grok 4.7): no blocking findings. A final sol check tightened one CHANGELOG sentence.