[egparedes]

Adopt Google Python Style, Conventional Commits, and cleanporter

What

Adopts the repository conventions end to end and reorganizes the test
suite to match them. Four phases, 17 commits, each green on make check:

1. Conventional Commits — commit-msg hook
   (compilerla/conventional-pre-commit v4.4.0, the 11 standard types,
   old areas become scopes), AGENTS.md rule with the old→new mapping,
   and a PR commit-subject check in CI.
2. Test-suite reorganization — the mixed files split by topic
   (inspectability, closures-and-helpers, error paths dissolved into
   per-module suites), the perf/fuzz/differential suites move into
   their level folders, every reference updated. Zero tests lost:
   956 → 970 collected (the layering non-blindness tests are new),
   862 → 876 passed, coverage 96% throughout.
3. Google Python Style — ruff gains D (pydocstyle google
   convention), N, ICN, TID252 and five zero-violation guards; all 251
   relative imports switch to absolute (ban-relative-imports = "all");
   62 docstring findings fixed by hand or autofix; one import per line
   per §3.13. Single documented deviation: 100-column lines.
4. cleanporter for §2.2 module-only imports — locked dev dependency
   (upstream v0.6.0, which carries the PEP 695 type-alias fix this work
   prompted: fix: bind PEP 695 type aliases as objects in the first-party layer cleanporter#10), fixer pre-commit hook,
   exit-code-robust make fmt, and the make lint gate with
   treat_unresolved_as_error. The tree is module-only everywhere
   except six documented inline suppressions (load-bearing re-exports
   and the optimize name collision) and the config-skipped golden test
   data.

Review guide

The commits are small and scoped; the two big mechanical ones are
style: rewrite src imports to module-only form and
style: rewrite test and script imports to module-only form
(review with git diff -M).

Verification

• make check: 970 collected, 876 passed, 94 skipped, 96% coverage
  (gate 95%)
• make lint: ruff + the cleanporter gate — 84 files, 0 violations,
  0 unresolved; proven to fail on injected violations
• make fuzz (5 seeds), make backends (93 passed), make lowest,
  make smoke, make docs: green
• Golden snapshots byte-identical throughout —
  git diff main...HEAD --stat -- tests/golden/cases is empty; no
  --update-golden was ever run
• Behavior-neutrality of the import rewrite audited at the AST level:
  only bare→module-qualified moves, the declared *_mod alias renames,
  and the six suppressions
• cleanporter's "import roots nest" stdout warnings on make lint are
  informational by design (inferred-root ranking), exit 0

Notes

• The cleanporter dependency pins the upstream v0.6.0 tag; the local
  stand-in that drove development is gone.
• Deferred follow-ups (tracked): emitter-side §2.2 compliance for
  generated code, the optimize-package cse/dce rename,
  tests/unit/test_builtins.py naming.