feat: add Kotlin universal evidence adapter

[forhappy]

Summary

• add the Kotlin universal evidence adapter for packages, imports, declarations, constructors, functions, extension functions, properties, annotations, generics, nullability, and named/default arguments
• add Kotlin-specific deterministic resolution, including member-vs-extension precedence and named/default/vararg overload applicability
• require exact compiler/SCIP evidence for Java/Kotlin cross-language calls
• convert Spring Kotlin to the universal spring-kotlin framework pack and remove its replaced legacy publisher
• harden update builds so fact-neutral edits skip project-wide resolution, refresh only file envelopes, point-update immutable store values, reuse unchanged graph JSON records, and amortize bounded store GC
• add fixtures, conformance coverage, pinned-corpus qualification results, and contract documentation

Impact

Kotlin code graphs now use the universal evidence pipeline and Spring Kotlin participates through a universal framework pack. The adapter remains UniversalCandidate: the independent 2,000-accepted-relationship precision/recall audit is not complete, and the pinned Spring Framework corpus still shows a cold-build performance regression.

The incremental fix reduces the 388-file Spring Kotlin one-file restoration from the previous 30.13-second observation to 1.01 seconds with zero graph assembly and zero new immutable objects. Release-build checks also covered ripgrep/Rust, go-git/Go, NestJS/TypeScript, the FastAPI production package/Python, and ASP.NET Core Http.Extensions/C#; all edited runs used the fact-neutral path and published valid SQLite-backed graphs. Detailed observations are recorded in PERFORMANCE.md and docs/implementation/kotlin-universal-qualification.md.

Root cause

compass update was excluded from the existing fact-neutral path. Kotlin also encoded whole-file hashes, ranges, and range-derived internal IDs into normalized extraction facts, so harmless EOF edits appeared semantic. Once resolution was skipped, publication still rebuilt broad store indexes and ran full graph-store garbage collection on ordinary edits. The fix makes the proof update-aware, preserves exact anchors and unrelated partial diagnostics, validates secondary-index invariants before point updates, and falls back to complete publication whenever the proof fails.

Validation

• cargo fmt --all -- --check
• cargo clippy -p compass-graph -p compass-core --lib --bins --all-features --locked -- -D warnings
• cargo test -p compass-core --lib --locked (76 passed)
• cargo test -p compass-graph --test store_snapshot --locked (20 passed)
• cargo test -p compass-languages --test kotlin_universal_conformance --locked
• cargo test -p compass-resolve --test universal_resolution kotlin --locked
• ./scripts/qualify_code_graph_v1.sh --fixtures-only (clean/warm/forced/incremental-restored/alternate byte equality)
• sh scripts/check_product_boundary.sh
• release-build real-repository incremental checks for Kotlin, Rust, Go, TypeScript, Python, and C#

The broader compass-graph --lib --tests run still exposes the pre-existing Markdown URI assertion in markdown_identity; the focused graph snapshot suite and the complete code-graph v1 qualification gate pass.