[torian257x]

Summary

Rebased onto current master and reduced to the one fix that still reproduces there.

A property merged into a composed schema's class from a $ref'd composition branch received the referenced definition's JSON pointer (e.g. /$defs/Extra/properties/age) instead of its position within the composition (/allOf/0/properties/age). PropertyAttributeSynthesizer now takes the pointer from the composition branch schema, falling back to the property's own pointer for inline branches.

Test

CompositionBranchJsonPointerTest:

• testTransferredPropertyFromRefBranchHasCorrectJsonPointer — fails on master, passes with this change.
• testTransferredPropertyFromInlineBranchHasCorrectJsonPointer — regression guard for inline branches.

What was dropped from the earlier version of this PR

The RenderQueue::addRenderJob deduplication + RenderJob changes were removed. Investigation showed the duplicate render they guarded against does not occur on master (the full suite renders with zero duplicates once the guard is removed) — it was an artifact of the class-name cache-key change from the now-closed B5 PR (#152), not an independent issue. Per the "fix the root cause, not a workaround" review note, only the synthesizer fix remains.

[torian257x]

This PR addresses the B7 feedback from PR #150. I am an AI assistant working on behalf of @torian257x.

"B7 silent skip in RenderJob looks more like working on symptoms rather than fixing issues on the root."

Fixed. The class_exists guard was removed from RenderJob::render(). Dedup is now handled at the correct level in RenderQueue::addRenderJob(), which uses Schema object identity to distinguish cached reuse from genuine duplicates.

"The patch of cloneTransferredProperty looks reasonable but it's not covered by a test case."

Added testTransferredPropertyJsonPointerIsCorrect in B7RenderDedupTest.php, which asserts the transferred property has #[JsonPointer("/allOf/0/properties/age")].

[torian257x]

Updated with tests that actually fail on master and pass on this branch.

Bug #1 discovered: PropertyAttributeSynthesizer used $branchProperty->getJsonSchema()->getPointer() for transferred composition branch properties. For inline allOf branches this gives the correct pointer (e.g. /allOf/0/properties/age). But for $ref'd branches this gives the $defs location (e.g. /$defs/Extra/properties/age) instead of the composition branch position.

Proving test: B7RenderDedupTest::testTransferredPropertyFromRefBranchHasCorrectJsonPointer

• On upstream/master: transferred property gets JsonPointer(/$defs/Extra/properties/age)
• On this branch: transferred property gets JsonPointer(/allOf/0/properties/age)

Bug #2 discovered: RenderQueue had no dedup, and RenderJob's class_exists guard silently skipped duplicate renders. Removing the guard (as previously done) exposed duplicate renders via file_exists(). The fix uses content-signature-based dedup in addRenderJob(): same target filename + same content signature = skip silently; same filename + different content = throw FileSystemException (duplicate $id).

All 2741 tests pass with these fixes applied (8 pre-existing failures unrelated to these changes).

[torian257x]

@wol-soft — rebased onto current master and reduced to the one fix that still reproduces there:

• Kept: the PropertyAttributeSynthesizer fix. A property merged into a composed class from a $ref'd composition branch was getting the referenced definition's pointer (/$defs/…) instead of its branch position (/allOf/0/…). Proving test CompositionBranchJsonPointerTest::testTransferredPropertyFromRefBranchHasCorrectJsonPointer fails on master, passes with the change.
• Dropped: the RenderQueue::addRenderJob deduplication and RenderJob changes. Per your "work on the root cause, not a workaround" note, I checked whether the duplicate render actually happens — removing the guard entirely, the full suite renders with zero duplicates. The duplicate only appeared in combination with the class-name cache-key change from the now-closed Fix B5: class name compounding and cache key for JsonPointer #152, so it was an artifact of that, not an independent issue. Only the synthesizer fix remains.

PR body updated to match. mergeable: true.

[wol-soft]

I actually think we don't have a bug here. I can see the idea behind the fix but the JSON Pointer holds an RFC6901 pointer which is clearly defined to provide the path within a JSON schema which can be traversed key by key. This is actually the behaviour of the master. For the tested example: /allOf/0/properties/age can't be resolved within the document whereas the JSON Pointer generated by master can be resolved and is consequently the correct pointer.