validation-conformance

validation-conformance

A behavioral cross-port corpus for generated input validation. The metamodel expresses field-input constraints two ways — field attrs (@required, @maxLength) and explicit validator.* subtypes (validator.{length,regex,numeric,array} with @min / @max / @pattern). Each port's codegen should enforce them in its generated input-validation artifact (TS Zod InsertSchema, C# DataAnnotations, Python Pydantic Field, Java/Kotlin jakarta.validation). This corpus pins the canonical accept/reject behavior so the ports cannot silently drift.

Shape

meta.json    # one entity `Account` (package acme::auth) exercising each constraint once
cases.json   # [{ name, payload, expectValid }] — single-source boolean verdicts
README.md

meta.json Account fields, one constraint each:

field	constraint
id	field.long + identity.primary @generation=increment (auto, excluded from insert)
name	field.string @required + @maxLength: 10
code	field.string + validator.length @min=3 + validator.regex @pattern="[A-Z]+" (un-anchored on purpose — see the full-match pin below)
score	field.int + validator.numeric @min=0 @max=100
tags	field.string isArray + validator.array @min=1 @max=3
label	field.string + @maxLength: 8 + validator.length @max=4 (both length bounds on one field — see the strictest-wins pin below)
note	field.string @required + validator.length @min=0 (an explicit opt-out of the FR-036 Pin 1 implicit floor — see the authority pin below)

Behavioral contract — payload → boolean verdict

Each port's runner builds the generated validation artifact for Account, runs each cases.json payload through it, maps the native result to a boolean valid, and asserts valid === expectValid. Verdict normalization per port: Zod safeParse().success; jakarta validator.validate(dto).isEmpty(); Pydantic construct-or-ValidationError; .NET Validator.TryValidateObject. The booleans are single-source — all ports assert the same expectValid.

Single-violation design

Beyond the valid-baseline (everything satisfied → true), every other case violates exactly one constraint while keeping everything else valid. That makes a simple boolean verdict gate each constraint individually: if a port fails to enforce constraint X, its violates-X payload is wrongly accepted and only that case fails — pointing straight at the unenforced constraint.

Cases: valid-baseline (true), then name-missing, name-too-long, code-too-short, code-pattern-mismatch, score-below-min, score-above-max, tags-empty, tags-too-many (all false). Every payload also carries a valid note ("ok") so these cases keep violating exactly one constraint (single- violation design, below) — note's own constraint is exercised by the note-* pin cases.

FR-036 pin cases (the two semantic pins + the length-precedence pin):

• name-empty ("" → false) and name-whitespace (" " → true) pin required string = non-empty (FR-036 Ruling 1): a @required string rejects null/"" but accepts whitespace-only. No port may trim (@NotBlank, [Required]-default) or emit nothing for a required string.
• pattern-unanchored (code="xxABCyy" → false) pins validator.regex @pattern = full-match (FR-036 Ruling 2): the whole value must match the authored (un-anchored) [A-Z]+. Ports whose regex primitive searches (JS RegExp.test, Pydantic pattern=) must anchor as ^(?:…)$; matches() (jakarta) / [RegularExpression] (.NET) are already full-match.
• both-length-ok (label="1234" → true) and both-length-bounds (label="12345" → false) pin @maxLength × validator.length @max precedence = strictest-wins: effective max = min(@maxLength, validator.max) (here min(8, 4) = 4), so a 5-char value is rejected even though @maxLength alone (8) would admit it.
• note-empty-allowed (note="" → true) and note-missing (note absent → false) pin an explicitly authored validator.length @min is AUTHORITATIVE over the FR-036 Pin 1 implicit non-empty floor (#224 / ADR-0044): note is @required with an explicit validator.length @min: 0, so the empty string is accepted (the opt-out) while the field must still be provided — @required's NOT NULL half is untouched by the opt-out, so an absent note is still rejected. No port may apply the Pin 1 floor (min 1) when an author has explicitly said @min: 0.

field.uri / field.inet strict contract (#234, ADR-0037 — Contract A)

Account carries two OPTIONAL fields — website (field.uri) and sourceIp (field.inet) — pinning the canonical STRICT accept/reject set. Optional so every pre-#234 case (which omits them) is untouched.

• field.uri = an absolute URI carrying a scheme (WHATWG / Zod .url() / Pydantic AnyUrl-aligned). ACCEPT https://a.com, ftp://…, mailto:a@b, urn:isbn:…, and leading/trailing-whitespace-padded (WHATWG trims). REJECT a scheme-less/relative reference (example.com, /path), garbage (not a url), and an empty authority (http://).
• field.inet = an IPv4 or IPv6 LITERAL only — no hostnames (never a DNS lookup), no CIDR (1.2.3.4/24), no padding, no out-of-range octet (256.1.1.1). ACCEPT 192.168.0.1, ::1, 2001:db8::1.

Deliberately UNPINNED gray zone (parser-specific, excluded from the probe set, like the corpus already does for regex/length): a raw space in a URI path (https://a.com/a b — WHATWG percent-encodes, java.net.URI throws); and value-normalization-on-accept differences — the corpus pins the accept/reject VERDICT, not the stored/echoed value. Notably an IPv4-mapped IPv6 literal (::ffff:1.2.3.4) is accepted by all five ports (pinned) but each normalizes it differently on read (Java collapses to the Inet4Address 1.2.3.4, C# keeps ::ffff:1.2.3.4, Python canonicalizes to ::ffff:102:304); likewise Pydantic appends a trailing slash to a URL. The corpus pins a FINITE probe set of verdicts, never total accept-set equality nor round-trip value equality.

Verdict normalization for native-typed fields. TS binds field.uri/field.inet to a plain string + a separable Zod check, so safeParse().success already fuses bind + validate. The other ports bind to a NATIVE type (Pydantic AnyUrl/ IPvAnyAddress; Java/Kotlin java.net.URI/InetAddress; C# System.Uri/ IPAddress) whose parser IS the validator — a malformed value fails to bind. So the rule is:

valid := (the payload binds to the port's generated wire-input artifact) AND (the artifact's declared validator reports no violations). A bind failure IS an invalid verdict — semantically honest: a request whose body fails to deserialize returns 400 exactly as a validation failure does.

Python fuses both (Pydantic construct-or-ValidationError); the Java/Kotlin/C# runners wrap the bind step so a native-parse failure maps to valid=false.

CI gate

All five port runners are wired into .github/workflows/conformance.yml (the non-Docker conformance job) — TS/C#/Java/Python under the conformance matrix, Kotlin under conformance-kotlin — asserting byte-identical boolean verdicts across all five generated validation artifacts.

That workflow runs on release tags and on workflow_dispatch, not on every PR (pull requests get the leak scan only, for hosted-minutes cost). Push-to-main coverage comes from the self-hosted local-ci.yml, and scripts/ci-local.sh reproduces the gate locally. See docs/CONFORMANCE.md.

Notes

• Cross-engine-safe regex. The code pattern [A-Z]+ is a plain ASCII character class — it behaves identically across the JS, Java, .NET, and Python regex engines. No PCRE-only constructs are used. It is deliberately left un-anchored in the metadata so the pattern-unanchored case can convict the full-match pin: a searching engine accepts "xxABCyy", a full-match engine rejects it. Every port must render full-match semantics (see the pin cases).
• Numeric field is int. jakarta @Min/@Max are long-valued; keeping score an int avoids decimal-string parsing nuance. A decimal-range case is a deferred enhancement.
• Deferred richer verdict. A per-field/per-constraint expectViolations: [{ field, kind }] is a deferred enhancement. The boolean verdict avoids per-port native-error mapping while still pinning each constraint via the single-violation design.