Do a bit of each, with rules. Retries detect flakiness and keep the pipeline moving but do not fix it: Playwright does not retry by default, and when retries are on it labels tests that failed and then passed as "flaky". Fowler's advice is to quarantine non-deterministic tests promptly so they stop blocking the build, then fix them, and to cap the quarantine (for example, a limit on how many tests can sit there, or on how long they can stay). Artsy's engineering team even fails the build when the quarantine exceeds its limit. Google warns that auto-quarantine can mask real race conditions and that marking tests flaky can encourage teams to ignore them. Delete tests whose value is already covered lower in the pyramid.

Sources: Playwright retries, Fowler: Eradicating Non-Determinism in Tests, Artsy Engineering, Google Testing Blog