Verify

NO COMPLETION CLAIM WITHOUT FRESH EVIDENCE

If the command did not run in this message, its result is not known.

The gate

1. Identify the command that proves the claim. Here: pnpm lint,
   pnpm type-check, pnpm test (or pnpm exec vitest run <spec> for one),
   pnpm test:e2e when behaviour moved or several files did, a live suite
   when its subject moved. Discover another repository's from its AGENTS.md,
   package.json or Makefile.
2. Run it, whole, on the exact tree you will push. git status clean of
   stray files first.
3. Read the output. Exit code, failure count, the runner's own totals. A
   background wrapper's exit code can be the echo's; a background vitest off a
   TTY prints totals only.
4. Judge. Output confirms the claim → make the claim, with the output.
   Output does not → state what actually happened, with the output.
5. UI surface (a component, a style, a layout, a role): drive it. The
   Playwright spec for the touched surface, or open the built app
   (pnpm desktop:build first; the electron e2e runs against out/). Green
   unit tests do not prove a render.

Where the evidence goes

• A goal brief: appended under ## Verification evidence. The verifier reads
  that section, not the conversation.
• A PR body: a "Verification" section with the commands and their last lines.
• A ledger job: the ledger_post or ledger_answer that reports the stage,
  in its body.
• A subagent's report: the command and its last line. A subagent that says
  "all green" with no output has not verified; check the VCS diff yourself.

Claims and what they need

Red flags

"should", "probably", "seems to", "Great!" before the output, "just this
once", a claim in a different wording so the rule does not apply.