Declaration Diagnostic Boundary

This document owns the cross-crate failure and diagnostic contract from discovery through declaration analysis. Public diagnostic meaning and stable codes remain owned by Diagnostics. Exact Rust variants and crate-local adapters belong to their defining source and rustdoc.

Production Path

DiscoveredUnit + canonical declaration surfaces
        |
        v
private declaration query
        |
        +--> accepted ReusableDeclarations
        +--> exact-current authored rejection
        `--> integrity unavailability
                    |
                    v
        closed program/unit analysis result
                    |
                    v
        session diagnostic and recovery evidence

nocter-compiler-computation is the production query owner. Eager declaration-lowering helpers exist only for focused tests and native composition tests; command, workspace, session, analysis, and LSP code cannot use them to establish a second failure order.

Failure Classes

Every failure has one owner and one class:

• source or syntax rejection: lexing or parsing owns the diagnostic; declaration analysis does not manufacture a second semantic error for the same malformed syntax;
• authored declaration rule: declaration lowering selects a stable rule and retains exact semantic subjects that can be projected to source;
• discovery contract failure: package, module, source membership, visibility, import, target, or toolchain input contradicts the frozen discovery snapshot;
• compiler integrity failure: an identity, arena, projection recipe, or completed earlier-stage contract is internally inconsistent.

Only the first two classes may produce public source diagnostics. Discovery contract and compiler integrity failures remain typed infrastructure failures; assigning them a language code would misrepresent a compiler defect as an authored mistake.

Rule Selection and Projection

The stage that detects an authored violation owns its rule identity and semantic subjects. nocter-diagnostics owns the phase-neutral envelope and rendering, but it cannot select a rule, repeat lookup, or infer a subject from display text.

Declaration lowering projects a complete authored report through the current source projection. The report is ordered canonically and duplicate-free before it leaves the stage. A missing subject or mismatched generation invalidates the whole projection; consumers cannot publish whichever prefix happened to project successfully.

SourceIndex locates an already selected semantic subject. It never decides that a declaration is invalid, changes diagnostic cardinality, or supplies a fallback rule.

Rejection and Recovery

An authored rejection is a first-class exact-current query value. It retains its complete diagnostic set and one branchable declaration recovery product. The top-level analysis query, not session, continues that recovery through the deepest body/name evidence permitted by the rejected declarations.

Recovery consumes admission facts selected during declaration validation. It cannot infer authorization from diagnostic codes, rerun a validator, or traverse rejected declarations to reconstruct the same decision. A declaration-only rejection cannot be converted into the input accepted by body analysis.

Session receives one closed success or rejection branch and preserves the complete diagnostic set in its failure envelope. CLI and LSP presentation read that same set; neither narrows it to one primary error and rebuilds the remainder.

Stage Ownership

The grammar conformance matrix assigns syntax and semantic-boundary cases to the narrowest owning stage. A declaration query cannot reject a valid intermediate program on behalf of checking, target validation, or a backend.

Required Invariants

• One authored violation has one rule owner and one source projection.
• Diagnostic rendering cannot affect compilation or recovery.
• Authored reports have deterministic ordering independent of traversal order.
• A lost source subject is an integrity failure, not permission to widen a range.
• Internal failure cannot become an authored diagnostic or an empty successful result.
• Session and feature code cannot restart declaration analysis.
• CLI and LSP consume the same closed diagnostic collection.