A refactor that changes existing tests

Tests and verification: one of the ways AI coding agents bloat software, from lessloc's pattern catalog. lessloc catches it as refactor-tests.

What makes a refactor that changes existing tests a bad practice?

The standard that calls it a bad practice: Fowler: refactoring preserves behavior.

lessloc catches it as refactor-tests.

The other patterns in Tests and verification:

Behavior added with no test changed in the same change: lessloc gives every session the rule: new behavior gets one focused test.
Tests that assert nothing, or assert a constant: lessloc catches it as empty-test.
Deleted, skipped or force-passed tests: lessloc catches it as weakened-tests.
Test-only branches in production code: lessloc catches it as test-logic.
Tests that mock the code they test: lessloc catches it as empty-test.
Production code that branches on a value from the tests: lessloc catches it as test-logic.
Stopping short: a requested function never written: lessloc catches it as unaddressed: names a request asks to exist ("add a --status flag", "implement export_csv()") are looked for in the final repository; one that appears nowhere holds the turn.
Stopping short: part of the problem left failing: lessloc catches it as unfinished: in a turn that changed code, the agent's last reply is read sentence by sentence, with headings and lead-ins speaking for what is under them.
Saying "done" without a passing run on the final state: lessloc catches it as tests-failed and unverified: the agent's own test runs, as its host reported them, are the evidence for the exact tree; a run that fails on the final state, or a change to how the tests run, holds the turn once, and a final state no run verified is written to the record; lessloc test, verify and ci record evidence too.