Bring the collector to 100% line coverage

[tschm]

Stacked on #25, which adds the coverage tooling this measures against. Merge #25 first and this retargets to main.

75.4% → 100% (637/845 → 845/845), in 153 new tests. No pragma: no cover and no omit rule — every line is genuinely executed.

module	before	after
github.py	61%	100%
__main__.py	30%	100%
localgit.py	84%	100%
metrics.py	90%	100%
config.py	98%	100%
state.py	96%	100%

The gap wasn't evenly spread, and neither is the value of closing it:

github.py — the HTTP layer had never run. Everywhere else the client is faked at _json, which is right for testing what the collector derives from a payload, but means the layer producing those payloads was untested: status handling, pagination, rate-limit bookkeeping. Those now run through an httpx.MockTransport, so the client, request building and response handling stay real and only the socket is replaced. That reaches the states a mixed fleet actually contains — a protection endpoint that 404s for two different reasons, an artifact GitHub already deleted, a feed longer than the page cap.

__main__.py — the layer that decides whether a failure degrades or takes the board down. Three properties now pinned: a raising refresh is recorded rather than propagated; a failure preserves last_success (zeroing it would say "never refreshed" about a board that refreshed a minute ago); and the port does not open until both halves are seeded. main() is driven by making the signal handler fire on registration — a second test schedules it instead, so the wait loop actually runs.

localgit.py — the branches a healthy checkout never takes, which is most of the error handling: git missing from PATH, a call that times out, a clone with no upstream, a tracked file deleted but not staged, a broken symlink, a hand-mangled template pointer, a repo with no commits at all.

metrics.py — one fixture with every optional field populated rather than twelve near-identical tests. The risk is uniform (a guard whose body never executed) and splitting it would obscure that.

Two things the writing turned up

• Only three Config fields re-read the environment per instance; the rest bind at import. A test that monkeypatches JQ_GITHUB_INTERVAL and calls Config() silently reads the value from process start — mine did, and failed. The new reload fixture documents that and restores the module afterwards, which the pre-existing hand-rolled reload didn't.
• runpy warns when re-executing an already-imported __main__ — fairly, since that's exactly what's happening. The module is dropped from sys.modules first rather than the warning suppressed.

Verified stable

Five consecutive runs and three randomized-order runs (pytest-randomly), 237 tests passing at 100% every time. Order-independence matters here because the suite reloads a module and starts threads and timers.