test: raise coverage above 90% (currently 32%)

[radu-mocanu]

Context

sonarcloud was wired up in #206. current overall coverage is ~32% (1131 statements, 768 uncovered). target is >90%.

only one test file exists today: tests/test_token_refresh.py. everything else is untested.

Per-module breakdown

module	stmts	miss	cover
_cli/_runtime/_runtime.py	414	299	27.78%
_cli/_runtime/_session.py	213	172	19.25%
_cli/_runtime/_tracer.py	126	110	12.70%
_cli/_runtime/_factory.py	65	44	32.31%
_cli/_runtime/_token_refresh.py	152	48	68.42%
_cli/_runtime/_context.py	16	6	62.50%
_cli/_runtime/__init__.py	8	3	62.50%
_cli/_utils/_config.py	73	49	32.88%
_cli/cli_new.py	47	36	23.40%
middlewares.py	4	1	75.00%

Suggested follow-up PRs

• cli_new.py and _utils/_config.py — smallest surface, mostly file IO and templating, easy wins
• _runtime/_factory.py — pure wiring of register_runtime_factory and the factory class
• _runtime/_context.py + _runtime/__init__.py — small helpers, finish them off
• _runtime/_token_refresh.py — already 68%, push remaining branches
• _runtime/_session.py — needs an in-memory MCP session double; biggest mocking effort
• _runtime/_runtime.py — the bulk; depends on session test infrastructure
• _runtime/_tracer.py — OTEL span assertions

Notes

• sonarcloud's new-code gate (default 80%) already blocks regressions on incoming PRs, so this can land incrementally without a big-bang test PR.
• tracer + runtime tests will likely need an anyio/mcp.shared.memory based in-process MCP server fixture.