tri catalog health: catalog Broken/Warn split by fixture and parse class, decided in t27 #6436

Refs #6434 (epic #6063)

Problem

The overnight t27b loop reads the SPECS catalog's Broken/Warn counts by hand or with a throwaway script. The raw counts mix negative fixtures (written to be rejected; bootstrap/tests/fixtures/**, neg_*, damage_*, eof_*) with real breakage, and parse failures (no AST nodes) with backend-only losses. There is no way to see progress between two catalog builds, and nothing says how stale the catalog's compiler is.

Change

• specs/tri/catalog/health.t27: the fixture markers, the bucket rules and the change rules between two builds, with tests; gen/c/tri/catalog/health.c by t27c gen-c (L2).
• tri catalog health: reads https://t27.ai/t27/manifest.json (or --manifest FILE), prints ok/warn/fail, the compiler commit the catalog was built from and how many commits that is behind origin/master, broken counts by repo and by bucket; --repo X, --list, --json, --diff OLD.json. Python is I/O only; every decision goes through the generated C via ctypes.
• CI: a gate test in loop-tools-gate (spec tests from the generated C, fixture manifest, mutation control).

tri t27b gaps is not added: tri t27b next already prints the sole/first blocker counts among specs where the reference passes and t27b is blocked.

Criteria

• python3 scripts/ci/test_the_catalog_health_is_decided_in_t27.py passes, and its mutation control fails a reversed rule.
• tri catalog health on the live catalog prints the totals and buckets.

Boundary

specs/tri/catalog/, gen/c/tri/catalog/, scripts/tri_loop/catalog.py, scripts/ci/test_the_catalog_health_is_decided_in_t27.py, .github/workflows/loop-tools-gate.yml, docs/now/

🤖 Generated with Claude Code