Replay Diagnostics Must Survive Until the Verdict

Scope: Engineering or methodology note based on a September 17, 2026 snapshot. Source-level behavior is not a public availability or universal compatibility claim. Any scenarios are illustrative, not customer outcomes.

A browser can display a convincing page while the replay engine records a problem underneath it. A request may be denied, a sequence exhausted, or the serving lifecycle interrupted. The visible page is not enough to decide which of those events matters to the comparison.

Kalu keeps replay diagnostics and lifecycle outcomes separate from a simple “server is running” signal. Product composition has to bind the relevant diagnostics into the measurement rather than ignore them after startup.

Startup is only the beginning

A loopback engine can bind successfully and encounter a runtime failure later. Conversely, a request-local denial may need to be reported without being confused with a terminal server failure. Those events require different handling.

The current replay handle exposes a lifecycle result when closure completes, including the primary runtime failure and any cleanup failure under its supported contract. Request-local denials remain in diagnostics rather than pretending they are all the same event.

Observe the diagnostics that affect the claim

In KodeBack’s cassette-parity path, relevant replay diagnostics raised during donor observation are bound into re-observation of the affected states. Diagnostics that continue changing can refuse the run. The comparison should not freeze a convenient early report while the underlying evidence is still unstable.

A synthetic example is a page that renders its heading before requesting a required late asset. The heading alone may look correct. The later denial can still matter to the declared contract and must not disappear from the final interpretation.

Keep the scope explicit

Not every diagnostic necessarily proves a candidate regression. Some identify missing evidence or unsupported measurement. The report needs to preserve that category so the operator can decide whether to recapture, narrow the scope, or investigate the candidate.

This article describes current source-level reporting and composition, not proof that every browser failure is observable. Coverage limits remain part of the result.

For an engineering review, ask what happened after the server announced its URL and before the verdict was accepted. A trustworthy replay story includes that interval. Otherwise, the product may be measuring the moment the page looked healthy rather than the journey it claimed to observe.