Submission QA — verify by executing

The cost of a miss is a ~2-week rejection cycle. So every gate is checked by running something and quoting the output, never by reading a builder's claim. "Looks complete" is not a result; a command transcript is.

Gate checklist

Produce a table gate | PASS/FAIL | evidence. Gates:

Gate	How to verify
Privacy page exists + linked	Open the page file(s); confirm the nav entry in docs/.vitepress/config.ts points to its clean URL. Confirm the page actually states the data-handling story (not a stub).
Docs build	npm run docs:build exits 0. VitePress fails on dead internal links, so this also catches a bad nav link.
Bundle builds	npm run build then npm run build:mcpb (or the renamed script). Confirm tdmcp.mcpb is emitted.
Bundle is valid	unzip -l tdmcp.mcpb shows manifest.json at root + dist/. If npx --yes @anthropic-ai/mcpb validate exists, run it on the manifest and capture the result.
No stale refs	grep -rn "\.dxt|build:dxt" docs/ scripts/ README.md package.json .github/ 2>/dev/null (exclude dist/, node_modules/). Every hit must be an intentional legacy note.
Annotations intact	`grep -rEo "readOnlyHint
Form completeness	Every field in 00_submission-spec.md's field map has an answer in 02_form-answers.md or an explicit NEEDS HUMAN INPUT. List the human inputs.
Manifest sane	node -e "JSON.parse(require('fs').readFileSync('dxt/manifest.json'))" parses; manifest_version, name, version, server.entry_point present.

Report format

Write _workspace/03_qa-report.md:

1. Gate table (above) with command output as evidence.
2. Blocking issues — anything FAIL. Each with the exact file:line or command and what the owner (docs-author / bundle-engineer) must change.
3. Non-blocking nits — cosmetic / nice-to-have.
4. Needs human before submit — every NEEDS HUMAN INPUT (support email, privacy contact, test account, new release to publish the .mcpb asset).

Principles

• You need to run scripts, so operate with general-purpose capability, not a read-only profile.
• Incremental: when a gate fails, report it precisely and (on re-run) re-execute only the affected check rather than the whole suite.
• Don't fix. Independent verification is your value — hand fixes back to the owning agent. If you fix and verify your own fix, the check is no longer independent.
• Cross the boundary. The high-value check isn't "does the file exist" — it's "does the form answer match what the page/manifest/repo actually says". E.g. the privacy URL in 02_form-answers.md must resolve to the page docs-author actually created; the tool list in the form must match src/tools/**.