Always read @AGENTS.md

Never commit data captured from a live system

This overrides convenience, and it applies to every fixture, example, README sample, doc snippet, seed file, decoder test and debugging artifact you write.

Fixtures are invented, never exported. If you are reproducing a bug with data from a running deployment — production, staging, lab, demo, customer or partner — that data does not enter the repository. Construct an equivalent synthetic case.

Forbidden when real; use the reserved alternative:

The table is about data taken from a deployment. Public reference data is fine in fixtures and demos: coordinates of real facilities such as airports, airport codes, airline flight numbers, vendor product names and OUIs, public place names.

Removing the organization's name is not enough. A naming convention, a coordinate pair, a build number, a serial, or a distinctive fleet shape identifies an organization on its own. If data turns out to be real, regenerate the fixture — do not search-and-replace it, because replacement keeps the shape and the shape is the tell.

Four things that make this worse than it looks:

• Non-test source ships. Decoder tests, README.md samples and docs-site pages are published product.
• Downstream registries are immutable. crates.io, proxy.golang.org / sum.golang.org, npm, hex.pm, OCI registries and the docs site cannot be fixed by rewriting git history. Check before you publish.
• Commit identity is content. Author/committer email cannot be corrected without rewriting history. Verify git config user.email in every clone and worktree; a global identity survives a re-clone.
• Do not name the affected party in a commit message, branch name or PR title — including when removing their data. Describe the change by class.

Anchor every search pattern. An organization abbreviation is usually a substring of ordinary English; here an unanchored three-letter match returned ~20,000 lines against ~480 real ones (equal, manual, actual, virtual, toEqual, quality). Fed to a history-rewriting tool, that corrupts every commit at once, unreviewably. Prove a pattern does not over-match before running it.

If you find captured data already committed, map how far it spread — other fixtures, published packages, the docs site, release tags — before deleting anything. The deletion is the easy half.