refactor: split metrics.render and the D-ranked functions, pinned by a golden exposition (#39)

[tschm]

Fixes #39

Acceptance criterion (verbatim):

uvx radon cc jq_collector -nc reports no block worse than C, and render is below C. The existing 447 tests stay green, and the /metrics output is byte-identical for the same snapshot.

Route: the issue left the split open; the maintainer asked for it with fix #39. The approach: pin the output first, then refactor under that pin.

1. The pin: tests/test_metrics_golden.py + tests/golden/metrics.prom

• One fixture fleet that takes every branch of render:
  • a fully populated repo, with duplicate and inconclusive workflows, a draft and a red PR, and a merged PR reported twice
  • a remote-only GitLab repo with a nested namespace, unknown protection, alerts off and a cancelled CI run
  • an unprotected repo whose clone is on a feature branch and out of sync
  • a local-only detached clone with no tags
  • an excluded repo
• The test compares the whole exposition, # HELP/# TYPE lines included, as one string. It covers 222 lines and all 51 families.
• The golden file was generated from the code on main before any refactor. The module docstring gives the one-liner that regenerates it after a deliberate exposition change, so the diff becomes the review.

2. The refactor (no behaviour change)

Block	Before	After
metrics.render	F (50)	A (4)
repos.load	D (24)	C (14)
GitHub.latest_runs	D (22)	C (11)

• metrics.py: render now yields the fleet-wide families from _health, then fills a _Families holder repo by repo. The fillers are _add_identity, _add_protection, _add_alerts, _add_drift, _add_ci, _add_pulls, _add_merged, _add_local and _add_size, and in_exposition_order() yields them. The 44 family definitions were moved by script, not retyped, so no HELP string or label list could drift. The yield order is taken from the old yield from (...) tuple. Every comment moved with its code. The largest new block is _add_local at C (11).
• repos.py: file reading and validation move to _read_entries, and the "listed twice" message moves to _listed_twice. The error messages are unchanged.
• github.py: the feed scan becomes _newest_per_workflow (pure), and one backfill call becomes _newest_conclusive. The insertion order into newest, and so the returned order, is unchanged.

uvx radon cc jq_collector -nc now lists only C blocks, 11 in all, the worst at 18.

Checks (run locally from collector/, as CI does)

• ruff check / ruff format --check / mypy: pass
• 3.12: 451 passed (the 447 from the criterion, 3 doctests, and the golden test), 100.00% coverage, floor met
• 3.11: 451 passed, 100.00%
• Byte-identity, both directions:
  • with origin/main's metrics.py in a scratch copy, the golden test passes. So the golden file is the old output, and the new code reproduces it.
  • a one-character change to a HELP string ("Stash entries." → "Stash entries") fails it.

Not addressed

• The remaining C blocks: localgit.scan_repo at 18, and _folder and _entry in repos.py at 17. The criterion stops at C.
• A return annotation on render.

Closes #39

🤖 Generated with Claude Code