No monitor on sign_in_failed and Desktop RUM is dark, so auth regressions cannot alert

A regression that cost roughly 620 sign-ins a day ran from 8 August until it was found by hand two weeks later. Nothing fired, and with the current setup nothing could have.

## Two separate gaps

**1. There is no monitor on `sign_in_failed`.**

`comfy.desktop.auth.sign_in_failed` is explicitly allow-listed for mirroring to Datadog in `src/shared/datadogMirroredEvents.ts`, and the comment on that list says adding a name to it means someone wants a monitor on that event. No monitor was ever created. The event has been mirroring into a void.

**2. Desktop's Datadog RUM application is effectively empty.**

It carries roughly 23 sessions in 30 days, against ~34 million events for the Cloud frontend. This is the more dangerous of the two, because it fails quietly in the wrong direction: anyone investigating a Desktop auth problem in Datadog gets a handful of rows back and concludes things are fine. Every number in the 8 August analysis had to come from PostHog for this reason.

## Asks

1. Create a Datadog monitor on `comfy.desktop.auth.sign_in_failed` with a rate-based threshold, not a count-based one, so it survives volume swings. The event-level failure rate sat at a stable 6% before 8 August and a stable 10.2% after, so a rate alert with a modest tolerance would have fired within a day.
2. Work out why Desktop RUM is dark and either fix it or write down that it is not a usable data source, so the next person does not draw a false conclusion from it. `sessionSampleRate` is set to 100 in `rendererBootstrap.ts`, so the sampling config is not the obvious cause.
3. Consider a per-person completion alert rather than an event-level failure alert. The failure rate moved 6% -> 10% across the regression, but per-person completion moved 70% -> 42%. The per-person metric is roughly seven times more sensitive here, because people retry and retries flatter the per-click rate.

Point 3 is the one that generalises. Failure rate per click is the metric we happen to have, but it consistently understates user-visible damage on any flow where retrying is cheap.