Export floor and breadcrumbs

Every log record, from Rust or from the page, passes through one floor (WARN by default, set with export_floor).

• A record at or above the floor is exported.
• A record below it is withheld in a short ring (12 records by default, set with breadcrumbs), and sent only if an ERROR follows while it is still there.
• It then goes as one extra breadcrumbs record, emitted just before the fault, carrying lines like:

-0.412s INFO webview: listing editor opened

Each line is the time before the fault, the level, the target (webview for the page's records) and the message.

The rules

• A WARN gets no trail. An expected condition is not something anyone opens a stack trace for.
• A trail is spent once it is sent. A second error a moment later does not resend the first one's context.
• With nothing withheld, there is no trail. It would duplicate records already sent.
• A withheld record is never exported on its own later. It rides with a fault or it never leaves.
• Events inside a span are held to the same floor. tracing-opentelemetry turns them into span events, and without the floor every debug! in a command would reach the collector through the trace instead.
• One WARN per launch names the service, version, environment, endpoint and floor. At the default floor an INFO would be withheld, and a quiet build could not tell you why it is quiet.

One ring serves both halves of the app, which is why the page's records go through the Rust side rather than through a second logger: the trail of an error in a Rust command includes what the page was doing just before it.

Designing for it

• Log at INFO and DEBUG freely. It costs nothing unless something fails, and then it is the context the error needed.
• Name your crates in filter. An event the filter drops never reaches the ring, so it can never be a breadcrumb. info,my_app=debug keeps your own DEBUG context.
• Use WARN for conditions an operator should see without a failure. A WARN carries no breadcrumb trail.
• breadcrumbs(0) turns the trail off. Records below the floor are then simply dropped.

Why

The design comes from a mobile client in this organisation that measured its own stream: 96.5% of what it shipped was debug and info chatter, one record per navigation and per state change. That costs the radio on every interaction and buys an operator nothing. A floor alone would cut that, and also cut the context an error needs to be understood. With breadcrumbs, the context survives the cut, and volume scales with faults rather than with use.