Add Step 3c's safety net: for every inbound golden fixture, run the raw
payload through both pipelines —
legacy: Inbound.Settings.fromJson(protocol, raw.settings).toJson()
zod: InboundSettingsSchema.parse(raw).settings
— canonicalize each (recursively sort keys, drop empty arrays / null /
undefined), and assert byte-equality. This locks the wire shape across the
upcoming class-to-pure-function extraction in Step 3d. Any normalization
drift introduced by the rewrite trips an assertion here before it can
reach users.
Two ergonomic wrinkles handled inline:
- The legacy class lumps hysteria + hysteria2 onto a single
HysteriaSettings (no hysteria2 case in the dispatch table); the test
routes hysteria2 fixtures through the HYSTERIA branch.
- Empty arrays in Zod's output (e.g. fallbacks: [] from a .default([]))
are treated as equivalent to the legacy class's omit-when-empty
behavior. Same wire state, different syntactic surface.
All 26 tests across 4 test files pass on first run.
