The false green: a cached "connected" cleared a live health check before any model ran

The cache was not lying. It returned the last known connector state, which is what a cache is for, and the author kept it on purpose: last known state is still useful during an incident, and the failed live path was the trigger condition rather than the bug ^[9]. The defect sat one layer up, in a health gate that accepted an answer to "what was last known?" as an answer to "is this pipeline healthy right now?" ^[18]. Cached evidence was allowed to support a present-tense verdict ^[10].

The shape of the evidence at the moment of that verdict repays a close reading. Seven fields went onto the pre-fix decision span ^[15], and taken one at a time almost all of them look reasonable: setup_state of connected, update_state of on_schedule, and an mcp_ok of false that is exactly what you would expect while telemetry is down ^[5]. The contradiction exists only as a join. A health_verdict of healthy sat next to an evidence.live of false, which means liveness was never a precondition of the gate in the first place ^[16].

That is also why ordinary monitoring finds nothing here. The /api/investigate call completed and the semantic failure raised no exception ^[6]. What made it legible was putting the healthy-looking values, their non-live provenance and the final verdict into a single trace, which changes the debugging question from why the endpoint says Healthy to why non-live evidence is permitted to support Healthy ^[17].

The repair is a third verdict rather than a pessimistic one. When connector health cannot be established, the app reports Needs review and names connector health as the likely issue; when the connector is treated as healthy while destination data is stale, it names upstream source freshness instead ^[8]. A binary of healthy and broken cannot carry that difference, so any tool with two states will eventually spend an outage investigating the wrong end of the pipe.

Worth stating the scope plainly. This is one developer's writeup, preserved as three intentional repository states and sequenced reproduce, instrument, inspect, diagnose, fix, verify, submitted to a DEV bug-smash contest powered by Sentry ^[13][14]. There is no outage cost attached, and no evidence about how common the pattern is elsewhere. What it does supply is a named failure mode with a concrete signature: a stale-but-plausible payload reaching a verdict path that never checked whether the payload described now.