aft_inspect reports "authoritative" 0 errors for a file outside project_root that has a type error #559

Summary

aft_inspect scoped to a file outside project_root reports FRESH and diagnostics: authoritative results for 1 of 1 scoped file with 0 errors, even when that file has a type error. The same file inside project_root reports the error. Per the reporting convention, an unexamined file should be named as a gap, not reported as clean.

Repro (aft 0.60.0, OpenCode 1 host, TypeScript repo)

The same one-line file, export const broken: number = "not a number";, in three places. project_root is the repo's main checkout.

1. Inside project_root, aft_inspect({scope: "<root>/packages/x/src/probe.ts", sections: "diagnostics"}):

   FRESH
   scope: 1 root, 1 file
   diagnostics: authoritative results for 1 of 1 scoped file
   diagnostics: 1 errors, 0 warnings, 0 info, 0 hints
   - packages/x/src/probe.ts:1:14 error Type 'string' is not assignable to type 'number'. [typescript]
   

2. In a linked git worktree of the same repository (outside project_root; node_modules installed, same tsconfig.json). Same call, run twice:

   FRESH
   scope: 1 root, 0 files (no analyzed files under this scope)
   diagnostics: authoritative results for 1 of 1 scoped file
   diagnostics: 0 errors, 0 warnings, 0 info, 0 hints
   

3. In an unrelated git repository outside project_root with its own tsconfig.json (strict: true): the same output as 2, plus TypeScript: server-managed SDK resolution (version not reported by AFT).

Expected

The file is reported as not examined, for example complete: false with skipped_files: [{file, reason: "outside_project_root"}]. That reason already appears in the incomplete_analysis_reason tests in crates/aft/src/commands/inspect.rs. Alternatively, diagnose it and report the error. Either way, never authoritative with 0 errors.

Where

The line is rendered from diagnostics.coverage.authoritative in crates/aft/src/commands/inspect.rs:3496 on main. I have not traced which producer counts an outside-root file as authoritative.

Why it matters

Parallel agent lanes run in linked worktrees. An agent that runs aft_inspect after an edit batch there gets a FRESH, authoritative, clean result, and type errors surface only at the next full typecheck.