Silent Zeros: When a Reporting Pipeline Says Success and Delivers Nothing

I repaired the merge so an empty input would not discard the other branch. On the verified scheduled run, the hourly reporting source received 96 rows instead of none.

I then wrote the fault-tolerant redesign: live staging imports, canonical text join keys, labelled import errors, freshness checks and a pipeline heartbeat that counts rows written. That architecture makes an empty run visible instead of presenting it as a plausible business zero. The evidence for this case establishes the merge repair and the redesign documents; it does not establish that every part of the redesign was deployed.

The conversions branch still returned no rows on the repair run. Restoring the reporting flow did not create missing source data or close that separate investigation.

I replaced the rejected phone-number matching route with click-based imports for inbound calls that carried a valid GCLID. I kept forms and calls in separate upload files, retained the lead-quality value tiers, and left the source API as the source of truth. The reporting sheet became a cache and audit copy.

The handover included the refresh schedule and the manual authorisation step needed to install its daily trigger. I documented which old upload routes should stay disabled.