New post: when did a test last fail on purpose?

[pftg]

New post: when did a test last fail on purpose?

Third post today, and deliberately NOT a third pass at the AI-content theme -
two in a row on that would trip the cross-post repetition gate. Different
cluster, different evidence, different question.

External hook: SQLite carried a data-race bug for sixteen years despite a test
suite larger than the database itself, and Tailscale's write-up contains the
line the whole post turns on - "so rare, the SQLite developers had to add code
to deliberately trigger it in their testing environments." They had to break it
on purpose before any test could see it.

First-hand evidence, all verified in-repo rather than recalled:

• The fault-injection run: eight realistic defects planted one at a time with
  predictions written down first. Three caught.
• Our link checker was skipping 133,874 of 149,516 links - checking about a
  tenth of the site and reporting the rest clean. Repointing it found five real
  defects that had been invisible, including a closing section offering a
  downloadable ROI calculator, itemised, "no email required, instant download",
  for a spreadsheet that never existed.
• The visual suite passing while comparing against nothing, and the detail that
  makes it worth publishing: three earlier attempts to break it came back green
  and the person doing it blamed their own injections. When a check is broken,
  the evidence that it is broken is indistinguishable from everything being
  fine.

Dedup: the archive's testing posts are practitioner how-tos (TDD tutorials,
suite performance, mocking). None argues that a passing suite is evidence about
the tests rather than the software, and none is written for someone who cannot
read the code. Four questions a non-technical founder can ask, each with a
checkable answer.

Gates: cadence quotas measured per H2 by script - FAILED first pass in two
sections (one with no sentence over 22 words, one with no single-sentence
paragraph), both corrected, all six now pass. Mannerism nouns were 10 against a
cap of 5 - "gate" is this post's subject matter, so five instances became
"check", leaving exactly 5 in the body. Banned words zero, em dashes zero.
bin/hugo-build green. marketing_copy_test 5/13/0. bin/rake test:links clean at
31,920 unique links, and the post is CONFIRMED present in the production build
at _dest/public-linkcheck rather than assumed.

Dated 2026-08-22, so live rather than scheduled.

1,017 words.

No independent verifier ran - agent spawning unavailable this session.

Co-Authored-By: Claude Opus 5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_016PUkwFTsiv7EB2DYKogbpg

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro Plus

Run ID: f344adf5-19a7-4075-9110-3c9b397d73b2

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}