fix(downgrader): keep binary content in 3.1 to 3.0

[dinwwwh]

The 3.1 → 3.0 downgrader no longer drops binary content. contentMediaType: image/png used to come out as {}, and only application/octet-stream became format: binary. Now raw and base64-encoded binary map to 3.0's format: binary and format: byte with type: string, following the spec's binary migration table. Upload schemas now reach 3.0 code generators as files or byte fields instead of any.

Fixes

3.1	3.0 before	3.0 after
{ contentMediaType: 'image/png' }	{}	{ type: 'string', format: 'binary' }
{ contentMediaType: 'application/octet-stream' }	{ format: 'binary' }	{ type: 'string', format: 'binary' }
{ contentEncoding: 'base64' }	{ format: 'byte' }	{ type: 'string', format: 'byte' }

• type: string is added only when type is missing, and an existing format is kept.
• Schemas whose type excludes string (e.g. object) get neither keyword, since content keywords only apply to strings.
• base64url and other encodings are still dropped: 3.0's byte is RFC 4648 section 4 base64 only.
• The README's schema table and "Removed" list now describe these keywords.

For reviewers

• Adding type: string to a schema with no type tightens validation, the one place this converter does. It is what the spec's table prescribes, and 3.0 tools ignore format without type.
• Every media type maps to binary, including text ones. So a JSON-in-a-string schema like { type: 'string', contentMediaType: 'application/json' } also becomes format: binary. Telling text and binary media types apart would need a guessed list, so this keeps one rule.

Testing

• Unit cases cover raw, encoded, nullable, multi-type, existing format, non-string type, and base64url.
• A new e2e test downgrades a raw image/png body and a base64 text/plain body and validates the result against the 3.0 schema.
• pnpm test (360 tests), pnpm lint and pnpm type:check pass. Existing snapshots are unchanged.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[pullfrog]

ℹ️ No critical issues — minor soundness suggestion inline.

Reviewed changes

• Content keyword conversion — new getContentFormat plus a reworked convertContentKeywords now map contentEncoding: base64 → format: byte and any string contentMediaType → format: binary, adding type: string when the source omits type and preserving an existing format or non-string type.
• Shared type helper — new hasType powers the content-keyword guard and the XML wrapped check (semantics unchanged).
• Tests — unit matrix for raw, encoded, nullable, multi-type, existing-format, non-string-type, and base64url inputs, plus a new e2e test that downgrades raw and base64 bodies and validates the result as 3.0.
• README — schema table and "Removed" list updated to describe the new mappings.

The spec's binary migration table (OAS 3.1.1 §4.4.2.1) matches the mappings, and pnpm test passes locally (360 tests).

  ^{｜ Fix all ➔ ｜ Fix 👍s ➔ ｜ View workflow run ｜ Using DeepSeek Flash (default — pick a model for stronger reviews) ｜ 𝕏}

[pullfrog]

ℹ️ Prior feedback addressed for const/enum; one residual case inline.

Reviewed changes

• Guarded the type: 'string' injection — new hasNonStringValue keeps the injected type: 'string' off schemas whose const or enum carries a non-string value, fixing the unsatisfiable outputs flagged last review. All-string enums still get type: 'string'.
• Tests — the content-keyword matrix now covers an all-string enum, a non-string const, and an enum with a non-string entry alongside the binary/byte formats.
• README — the contentEncoding: base64 row documents the new "every const/enum value is a string" condition, and the "Removed" list names a non-base64 contentEncoding with its contentMediaType.

  ^{｜ Fix all ➔ ｜ Fix 👍s ➔ ｜ View workflow run ｜ Using DeepSeek Flash (default — pick a model for stronger reviews) ｜ 𝕏}

[pullfrog]

✅ No new issues found.

Reviewed changes

• Skipped type: string beside type-constraining siblings — hasNonStringValue became hasTypeConstraint, now also true for $ref, allOf, anyOf, oneOf, and not, so the injected type: 'string' no longer sits beside a combinator/ref that pins the value type. format is still emitted at the parent level (non-validating in 3.0).
• Tests — exact unit assertions pin the allOf, nullable anyOf, and $ref cases; each would fail without the guard.
• README — the contentEncoding: base64 row now lists the $ref / combinator keywords alongside the non-string const / enum condition.

This closes the residual unsatisfiable-output class raised in both prior threads. pnpm test passes locally (365 tests) and the checked-out converter confirms the new outputs.

  ^{｜ Fix it ➔ ｜ View workflow run ｜ Using DeepSeek Flash (default — pick a model for stronger reviews) ｜ 𝕏}

[pullfrog]

✅ No new issues found.

Reviewed changes

Since the prior pullfrog review (664d285), commit 33b3d5d replaced the hasTypeConstraint guard with a single rule.

• Dropped the type-constraint guard — convertContentKeywords now injects type: string whenever the source omits type; the $ref / allOf / anyOf / oneOf / not and non-string const / enum checks are gone.
• Pruned the guard tests — the six rows pinning the skipped-injection cases (non-string const, mixed enum, all-string enum, allOf, nullable anyOf, $ref) were removed.
• Simplified the README row — the contentEncoding: base64 description now reads "plus type: string when type is missing".

This intentionally reintroduces the unsatisfiable-output behavior the two prior threads raised: { const: 42, contentEncoding: 'base64' }, { contentMediaType: 'image/png', enum: ['x', null] }, and { contentMediaType: 'image/png', allOf: [{ type: 'object' }] } again emit type: string beside a non-string constraint. The author documented the rationale in-thread (contradictory source intent, generators don't emit these shapes, shorter code), so no change is requested. The realistic nullable shape type: ['string', 'null'] still converts to { format: 'binary', nullable: true, type: 'string' }, and pnpm test passes locally (359 tests).

  ^{｜ Fix it ➔ ｜ View workflow run ｜ Using DeepSeek Flash (default — pick a model for stronger reviews) ｜ 𝕏}