fix(compilers/openapi)!: record $self on the document's source

[fuad-daoud]

Summary

• ir.SourceInfo gains SelfURI (json:"selfURI,omitempty"), and the OpenAPI compiler fills it from doc.GetSelf(). It sits on the source entry beside format, path and hash because $self is a fact about one input file, not about the document: several may be loaded, each may declare its own, and a document-level field could not say which file declared it.
• The URI is recorded at every supported minor, not gated to 3.2. The library parses the keyword at every 3.x minor and reports no finding for it, so a version gate would silently drop a value a document wrote. A 3.1 document that writes $self is the case that would have been dropped; the recorded value is the declared one either way.
• $self: "" and an unset $self are one state: an empty URI reference is not a base URI (RFC 3986 §5.1), it identifies nothing when documents are combined, and nothing downstream can act on the difference. That is why the field is a plain string with omitempty.
• Deliberately out of scope, stated here and in the field's GoDoc: resolving relative references through the URI (GitHub engine: multi-file specs — bounded load-and-bundle and a sources-based entry point #74 owns that wiring), and any diagnostic for a $self written by a dialect that does not define it — the analogous location check for allowReserved is openapi: allowReserved on a path or header parameter is kept without a warning #628, not this change. This change adds no diagnostic code.
• Nothing else moves: no openapi.Options, pass/, engine/, CLI or irverify change.
• D5 — IRVersion stays 0.6.0. docs/ir-design.md §2.1 moves the version for "a key renamed or removed, an encoding changed, or the meaning of an existing key changed"; a new optional member is none of those, and the one pure field addition in the history (Property.Constraints, ecd2e78) landed without one. Counter-reading, stated not hidden: ir/document.go warns that a shape change reaching main without a bump leaves a consumer pinned to the old version accepting a document it cannot read, and a 0.6.0 build handed a document carrying selfURI does fail to decode it — loudly, on the unknown member. This PR takes the §2.1 enumeration. If the batch owner rules the other way, the follow-up is mechanical and touches nothing else here: bump the constant to 0.7.0 with a log paragraph, re-anchor TestCompatibleVersion's three neighbour literals, run the -update sweep, and title the PR feat(ir)!: ….

Test plan

• go test ./ir -run TestSourceInfo — the wire shape: a declared URI under selfURI byte-exact beside format/path/hash, and no key at all when unset; each row also decoded back.
• go test ./compilers/openapi -run TestParse_RecordsSelfURI — five rows: 3.2 with $self, 3.2 without, 3.2 with $self: "", 3.1 with, 3.0 without; each asserting no error diagnostic, one source, and Sources[0].SelfURI.
• Corpus: testdata/conformance/openapi/self-uri.yaml (3.2.0, $self, info, paths: {}) with {"self-uri", assertSelfURI, nil} and its golden; go test ./compilers/openapi green, unwitnessed.golden.txt unchanged.
• Planted mutations: deleting SelfURI: doc.GetSelf() reddens TestParse_RecordsSelfURI/3.2_with_$self, /3.1_with_$self and TestConformance/self-uri; deleting $self: from the fixture reddens TestConformance/self-uri on the capability assertion and leaves the golden unrewritten — require.False(t, t.Failed()) guards the snapshot from being rewritten behind a regression — while a byte-deleted selfURI key is restored byte-identically by -update.
• make gate green on 804b9ae, with main merged in: coverage at exactly 100%, golangci-lint 0 issues. The two doc comments over the 100-word cap that failed the earlier CI run are rewritten.
• The 0.8.0 bump regenerates every golden; the only change in each is the irVersion line, and TestCompatibleVersion's neighbour rows are re-anchored.

Breaking

ir.IRVersion moves from 0.7.0 to 0.8.0. SourceInfo gained selfURI, and ir-design §2.1 moves the version on any change to a Document's JSON shape: a consumer built against 0.7.0 refuses the new member, so a document carrying it is unreadable there. The version history in docs/ir-design.md records the entry. #734 and #735 also claim 0.8.0, so whichever of the three merges later moves to the next version.