Language diagnostic drift — findings (2026)

Language diagnostic drift — findings (2026)

Probe charter

Determine whether the same logical mistake can surface under different IDs, severities, or messages across:

• vox-compiler / vox check
• crates/vox-lsp document validation
• vox-code-audit detectors
• Optional Web IR / emit validators

Evidence (2026-05-11)

• LSP maps typechecker diagnostics with source: Some("vox-lsp") while preserving string code from the compiler (vox-lsp/src/lib.rs).
• Phase plans mandate stable IDs and --for-llm JSON (Phase 2 plan); implementation completeness must be tracked separately.

Hypotheses

1. Source field drift: Consumers keyed on source may dedupe incorrectly between "vox" and "vox-lsp".
2. Detector overlap: TOESTUB / audit rules may duplicate compiler warnings with different naming until SSOT collapse completes.
3. HIR-only diagnostics: Extra HIR validation in LSP when include_hir may diverge from default CLI flags.

Recommended next steps

1. Add a small golden fixture set: one .vox file per canonical diagnostic; snapshot CLI JSON vs LSP payload.
2. Document precedence rules in vox-diagnostic-ux-ssot-2026.md once measurements exist.
3. Optionally unify source string or add explicit tool field in --for-llm output (future RFC).

Related

• vox-diagnostic-ux-ssot-2026.md