This directory holds canonical-text snapshots for each non-canonical face's output, plus a tiny harness for confirming they keep parsing.

For every example under examples/faces/:

1. Snapshot diff — the script runs the corresponding preview-* subcommand and diffs its stdout against a committed *.expected.txt here. If the transformer's output for the same source ever changes, the diff fails.
2. Round-trip parse — the example file is parsed via the normal pipeline (which auto-detects face from the pragma). If a transformer change ever produces canonical text the parser rejects, this catches it.
3. Canonical baseline — examples/faces/hello-canonical.affine is parsed directly to confirm the reference shape stays valid.

just build
just test-faces-record   # captures any missing snapshot, then diffs
git diff tests/faces/    # review what was captured
git add tests/faces/*.expected.txt
git commit

just test-faces           # diffs against committed snapshots; fails on drift

Edit a face transformer (e.g. lib/python_face.ml), then:

just test-faces-update    # overwrites the affected snapshot
git diff tests/faces/     # review the lowering change
git add tests/faces/*.expected.txt
git commit

The diff in the PR shows reviewers exactly how the canonical lowering changed, which is more useful than just "transformer modified".

The transformers are pure text-to-text. Bugs typically show up as drifted output rather than crashes — a missing comma, a wrong keyword swap, a broken indent rule. Snapshot diffs catch those instantly. Combined with the round-trip parse, this gives a regression net that:

• runs in seconds (no codegen, no wasm),
• has zero false positives (output is deterministic),
• doubles as a side-by-side reference for "different faces, same cube".