[harper-joseph]

The console last caught up at plugin v0.55.0. Plugin v0.56.0 and v0.57.0 added three prerender_ops series and reshaped both node-local passes, and the console read none of it.

It shipped invisible because the metric-coverage guard exempts metrics.changeProbe — it names its series with a probe_${series} template, which the guard's regex structurally cannot read. Green suite on both sides. One of the three missed series is the one the catalog says to alert on.

Change probe

series	what it means	why it needed a surface
probe_throttled	origin pushback — 429/502/503/504, connect/read timeouts	The only signal that the probe is loading an origin that cannot take it. The sweep answers by halving its own rate, so it covers less of the corpus per pass while the change rate, failure share and trigger count all keep exactly the shape they had.
probe_fresh	rows skipped because a baseline was younger than reprobeAfter	Explains why probed drops after a restart. Sustained, it means reprobeAfter sits too close to sweepInterval and real cadence is quietly two sweep intervals.
probe_unreadable	registry rows whose key would not decode	Those targets are never probed and appear in no other count.

• throttled gets its own alarm, a head pill, and a per-node column — the probe rate is agreed and held per node, so pushback is a fact about one node's conversation with the origin, and a cluster total names nobody to go and look at.
• fresh is counted against the rows a pass considered, not against its probes: a skipped URL was never attempted, so it sits outside probed entirely.
• unreadable is stated as a storage-layer escalation, because no setting on that page reaches those rows.
• The three ways a pass can stop early are no longer one Interrupted label. Standing down for a reseed is routine; giving up on a refusing origin means the slice was never covered and the counters are a partial count.

Discovered-target purge

• skipVisited, defaulted ON in the console although the plugin defaults it off. A plugin default that changed behaviour for existing callers would be the wrong kind of change; a console default is a suggestion to a human. Sent on the census as well as the purge, so the number an operator approves is the number that happens.
• visitedSkipped and errors are now terms in the "not reached" subtraction. Without them a pass that spared 40% of a prefix on purpose reported that 40% as keyspace it never got to — a completed pass rendered as an interrupted one. (unreadable is deliberately not a term: the walk skips those rows before anything reads their sitemapUrl, so they never entered discovered.)
• abortedOnErrors and the delete-failure samples, because a failed delete leaves a row in the corpus that appears in no render metric, no serve metric, and nowhere else.

Cluster merge

Counters that sum, sum. The pacing window is taken as the worst node's, never a mean — averaging one throttled node against three healthy ones answers "was the probe being held back" with a number that was true nowhere. Nodes that gave up on a refusing origin, or ran a different delete predicate over the same prefix, are named rather than folded into a total.

Guard

A second metric contract test covers the probe family via the catalog's own values list — the source the regex-based guard cannot reach. Verified by hand to fail when a series goes unread, and to pass again when restored.

The other dynamic families (sitemap_*, demand_*, queue_health gauges, invalidation_*) stay unguarded, and the test says so explicitly. They are legitimately unread from analytics — the queue gauges come from the overview endpoint, the demand series have no panel — so guarding them means a waiver list stating a decision per series. That belongs with whoever makes those decisions, not in a catch-up.

Not in this PR

changeProbeStatus() does not expose a rule's statusSignals, so the probe view's rules table cannot show which rules treat a 4xx as an observation rather than a failure. That needs a plugin field and therefore a plugin release train; the console's failure-share narrative stays correct without it. Worth a follow-up.

Verification

• packages/console: 210 pass, 0 fail (was 185 — 25 new tests)
• packages/plugin: 850 pass, 0 fail, untouched
• npm run lint and npm run format:check clean at the root

🤖 Generated with Claude Code

[gemini-code-assist]

Code Review

This pull request bumps the version of @harperfast/prerender-console to 0.9.0 and introduces several UI and aggregation improvements to the management console. Key additions include support for sparing bot-visited targets (skipVisited) during discovery purges, handling and displaying origin pushback/throttling, tracking skipped fresh rows from resumable sweeps, and surfacing unreadable registry rows. It also updates the aggregation logic to correctly merge these new metrics across cluster nodes and adds comprehensive test coverage for these features. I have no feedback to provide as there are no review comments.