Lite bands no collection freshness, so a server can go quiet and stay green

[erikdarlingdata]

Two things #2437's Lite half (#2451) surfaced and deliberately did not fix, because both are wider than a tooltip.

1. Lite bands no collection freshness at all, so a card can go quiet and stay green

Writing Lite's card tooltip meant reading every signal the card actually carries. It carries five metric rows and a status word, and none of them is freshness.

• LastCollectionDisplay renders HH:mm:ss in the plain ForegroundBrush — no band, no brush, no threshold. A timestamp from four hours ago looks exactly like one from four seconds ago.
• ServerSummaryItem.IsOnline comes from _serverManager.GetConnectionStatus(...) — a live connection check, which succeeds whether or not anything is being collected.
• HasCollectorErrors comes from ErroringCollectors > 0, which counts collectors with ConsecutiveErrors > 0. A collector that silently returns nothing, or one that has stopped being scheduled, increments nothing.

So the failure mode is: the connection is fine, no collector is erroring, and the store has had no new rows for hours. Lite's card shows a green "Online", a neutral border, and a stale timestamp nobody is asked to read. The Darling viewer bands exactly this and calls it Warning — collection stale (ServerHealthThresholds.StaleThreshold = 2x the fastest collector's cadence, OfflineThreshold = 15 minutes).

The tooltip added in #2451 does not paper over this, on purpose: it is built from the card's own rows precisely so it cannot assert a band the card does not show. Adding a "collection stale" clause to the tooltip while the row underneath stays uncoloured would have been the exact drift that property exists to prevent.

What would settle it: the thresholds already live once, in PerformanceMonitor.Common's ServerHealthThresholds, and Lite already has LastCollectionTime on the card. So the options are:

1. Band the Last Collect row from ServerHealthThresholds and let the tooltip name it, the way it names CPU and Blocking today. Smallest change, keeps the card as the single source, and the tooltip picks it up for free through the existing gate pattern.
2. Fold freshness into the status word, as the viewer does — which would make Lite's ServerCardStatus.CollectorErrors mean two things again, and is the shape Say why the card is amber, and give "+52 more" somewhere to go (#2424) #2429 spent four review rounds untangling. Probably wrong here.
3. Leave it, and accept that Lite reports collector errors but not collector silence.

The measurement that should decide: whether Lite installs actually see collection stop without a collector erroring. Lite's collection_log equivalent per server would show it directly — if a gap of more than a few cadences with zero ERROR rows is common, option 1 is not optional.

2. ServerConnection.DotStatus is a fourth copy of the same four-word ladder

Lite/Models/ServerConnection.cs:199 computes "Unknown" | "Online" | "Warning" | "Offline" from its own (IsOnline, HasCollectorErrors) pair, for the sidebar row's dot.

#2451 collapsed the Overview card's three readings of that pair into one ServerCardStatus, for the reason #2429 established: with one discriminant there is no flag combination left for the renderings to disagree about. DotStatus is a different type on a different surface and was out of that PR's lane, but it is now the remaining place the same words are derived independently — the sidebar and the card can say different things about the same server, and nothing would notice.

It also has no tooltip, which is the #2422 complaint one surface over.

Low priority on its own; worth doing at the same time as (1), since a freshness band would otherwise need writing twice.

[erikdarlingdata]

Part 1 is up as #2457: the Last Collect row is banded from the shared ServerHealthClassifier/ServerHealthThresholds (no new numbers in Lite — its fastest collector really does run at CollectorCadence), and the status word is deliberately left alone, which is option 2 being turned down with the reason written on StatusDisplay.

One thing the issue asked to be measured came out of Lite's own source rather than an install: RecordCollectorResult treats PERMISSIONS as not-a-failure and leaves ConsecutiveErrors alone, and IsCollectorPermissionRestricted returns out of RunCollectorAsync before LogCollectionAsync — so a permission-restricted collector writes no collection_log row at all and increments nothing. A login that loses rights across the board stops the store dead with ErroringCollectors at zero and every card green. The in-memory health map makes it worse across a restart.

Part 2 (ServerConnection.DotStatus) is split to #2458. The issue's reason for pairing them — "a freshness band would otherwise need writing twice" — did not survive: the band is written once, on ServerSummaryItem, and the sidebar dot is untouched. What does argue for splitting is that the type DotStatus should collapse onto is ServerCardStatus, which arrives in #2451 and is still open.

[erikdarlingdata]

Part 1 fixed on dev by #2457. Part 2 is #2458.

ServerHealthThresholds is shared, with no new numbers in Lite — it already referenced PerformanceMonitor.Common, so it cost one using. And the sharing is justified on merits rather than convenience: Lite's fastest configured collector in collection_schedule.json runs every minute, which is exactly the derivation CollectorCadence documents, so StaleThreshold is true for Lite on its own terms rather than borrowed.

Freshness deliberately stays OUT of the status word, and I think this is the right call even though it is not what the issue implied. Lite's word is a connection word and CollectorErrors already means one thing; folding freshness into it would make one amber word mean two unrelated failures — which is exactly the #2429/#2422 conflation, imported rather than avoided. So the Last Collect row bands itself instead (08:00:00 (stopped) in red) with a tooltip saying plainly that this is about collection stopping rather than about the server's metrics, and that the connection check can still be passing with no collector erroring.

A quiet server therefore still reads a green "Online" — correctly, because the connection check did succeed. The card now says two true things in two places instead of one ambiguous thing in one.

Worth recording how a defect got caught: the agent ran #2451's committed tests against this branch before CI did, and found its own CardBorderBrush edit had broken their landed source pin. CI then failed on exactly that one test and nothing else — so the local check and the remote one agreed, which is the useful signal.

Part 2 (ServerConnection.DotStatus) is #2458, and the reason I gave for pairing them did not survive contact: the band is written once and the sidebar dot is untouched by this change. What does argue for doing it is different — DotStatus should collapse onto the ServerCardStatus that arrived in #2451, so it is a consolidation rather than a second banding.