fix(compilers/openapi): keep a redeclaration's losing type

[fuad-daoud]

Closes #424. Not breaking.

When allOf branches declare the same field with incompatible types, the first declaration wins, a warning names both pointers, and the losing type was simply dropped — not written to Unmodeled.

Every other degradation in this compiler keeps what it could not model. That is the discipline: an Unmodeled entry with a Reason, so nothing reaches the IR in no form. This path was the exception, and it was silent to any consumer reading the document rather than the diagnostic stream.

102 occurrences in GitHub's published spec — e.g. webhook-fork/properties/forkee/allOf/1/properties/clone_url, where t/prim/url and t/prim/string disagree and the first wins. A consumer diffing two versions of that document sees "no change" across a release where the losing branch's type moved.

The whole ir.TypeRef is kept (so Nullable survives, not just the target ID), keyed by the redeclaration's own pointer so three conflicting branches keep three entries rather than the last overwriting the rest.

Scope note: the constraint half of the same diagnostic still discards its loser. That is left alone deliberately — #10 records the direction for it (intersect the bounds), and preserving the loser instead would settle a decision that already has one.

---

Stack 2 of 8. Base stack/1-detection — review and merge bottom-up. Every commit here passed make gate when it landed, and the full gate was re-run on the top of the stack. Run it as GOTOOLCHAIN=go1.26.3 make gate; this machine's Go 1.27 fails it for reasons unrelated to any change (#431).

🤖 Generated with Claude Code

https://claude.ai/code/session_01SYeBgDsskwnyitLgPGCPn1

---

Review round 2 (commit 2)

Addresses the review on this PR. Three findings shared one root — the discard was partial and the preservation gate was narrower than the drop — and are fixed together in reconcileProperty.

The discard was partial. After a type conflict fired, the fold ran to completion anyway, so {id: integer} allOf {id: string, maxLength: 10, default: "abc"} compiled to an integer carrying a string default and a string's maxLength, beside an entry saying the string declaration was discarded. Nothing downstream compares a Value's kind to its property's type. recordRedeclarationConflict now returns whether the type was discarded, and Default/Constraints/Examples are gated on it; Deprecation and XML are not shape bound and are adopted either way.

The gate was narrower than the drop. Preservation hung off typesConflict, which deliberately answers false for two composites of one kind, an unresolvable target, and the top type against anything. In each of those dst kept its type and src's vanished silently — #424's own failure. Preservation is now owed wherever a type is dropped; the diagnostic stays on the conflict predicate, since "dropped" and "contradictory" are different claims.

Nullability was neither reconciled nor preserved. typesConflict returns on a.Target == b.Target before Nullable is read, so {x: [string, "null"]} allOf {x: string} merged to nullable with no diagnostic and no entry — and swapping the branches gave the opposite answer. Targets that agree now intersect nullability. The order oracle could not have caught this: it never permutes sequences.

Also fixed: the type route's diagnostic now names the preservation, as every other preserving degradation here does; MergeProperty takes the declaration's position once, off p.Provenance, instead of as a second parameter equal to it by construction; keepLosingType guards an empty pointer or target, which would collapse the key to the bare prefix and let a second loser overwrite the first; diagnoseRedeclarationConflict is renamed for the recording it does and its comment corrected; the false universal about every degradation preserving what it could not model is deleted; ir-design.md gains the §4.8 entry and the §14 mapping key this key was already cited against; and the two tests stop re-implementing their helpers — one was failing on assert.JSONEq(want, "") rather than on the missing key.

Title shortened to 65 characters with the number, from 77.

Known gaps, deferred deliberately

Both are stated in code beside the relevant function and filed:

• conflicting-redeclaration keeps an IR TypeRef where ir-design 12 asks for the source construct #445 — the Unmodeled value is an IR TypeRef where ir-design §12 asks for the source construct. irverify's reference walk cannot see it dangle, and a losing branch carrying residue is preserved twice. Fixing it moves MergeProperty's signature and the golden, where this PR's scope is conflicting-redeclaration discards the losing type with no Unmodeled record #424.
• typesConflict reports a format-narrowed primitive as conflicting with its bare primitive #446 — typesConflict reads a format-narrowed primitive as conflicting with its bare primitive, so {string, format: uri} against {string} reports a degradation that did not happen — 102 times in the published GitHub spec. The predicate is upstream of this change, and moving it shifts goldens across the corpus.

[Wahbeh-Mohammad]

Two-pass review of this change (design/coherence, then adversarial), verified against the source at 8a42a7f. Inline comments carry the findings and suggested fixes. Three of the majors share one root: the "discard" is partial (annotations still fold) and the preservation gate (typesConflict) is narrower than the drop. They are cleanest fixed together in reconcileProperty.

One item with no line to anchor to:

Major — PR title over the cap. printf '%s (#436)' "<title>" | wc -c gives 77; the squashed subject cap is 72. Suggested: fix(compilers/openapi): keep a redeclaration's losing type (63 with the number).

Vetted and sound, for the record: site and layering (merge, not pass/); PreserveInto reuse and its empty-payload guard; entry on the merged property per §12.1; pointer in the key for order-invariance, RFC 6901-escaped so injective; MergeUnmodeled runs after and cannot clobber the new key; N-way conflicts yield N-1 entries; golden reddens on revert; the constraint-half carve-out is stated in code.

[Wahbeh-Mohammad]

Review done by me + fable