Motivation

Filtered<L, F> carries each filter's per-call decision from Layer::enabled to Layer::on_new_span/on_event via the thread-local FilterState bitmap. That side-channel assumes enabled() runs immediately before on_new_span/on_event for the same metadata. When a callsite caches Interest::always, the tracing macros skip enabled() entirely — so on_new_span reads whatever the previous enabled() call (for unrelated metadata) wrote, and the span is silently dropped from every per-layer-filtered layer.

The common production trigger is log::log_enabled! via tracing-log's LogTracer, which calls dispatch.enabled() with per-call dynamic metadata that bypasses callsite caching. Any crate that does this between two #[instrument]ed calls — sqlx's QueryLogger does after every query — leaves stale disabled bits in FilterState, and the next Interest::always span never reaches the OTel/fmt/etc. layer it was meant for.

Fixes: #2519
Refs: #3516, #2448, #2704

Solution

Registry::register_callsite now caps the per-layer-filter interest sum at Interest::sometimes. Callsites that every per-layer filter accepts get sometimes instead of always, so enabled() runs on every dispatch and FilterState is always fresh when did_enable reads it. Callsites any filter rejects are unaffected (never if all reject, sometimes if mixed — same as before).

Trade-off. This removes the Interest::always fast-path for callsites every per-layer filter accepts. Those callsites now pay one Filter::enabled() call per dispatch instead of zero — for the common filters (LevelFilter, Targets, EnvFilter with static directives) that's a level compare plus a target prefix match, plus one thread-local write in Filtered::enabled. Stacks without per-layer filters are entirely unaffected. The four *_interests_are_cached tests are updated to reflect the new invariant: rejected callsites still cache as never; accepted callsites are re-evaluated per dispatch.

Users are already paying a comparable cost via the "add a no-op layer" workaround (transact-rs/sqlx#3751), which makes rejected callsites sometimes instead. This PR is cheaper for the typical case where rejected callsites outnumber accepted ones, and it doesn't depend on a follow-up event being dispatched to reset the bits.

Scope. This fixes the bare-enabled() trigger of #3516 — LogTracer, the enabled!/event_enabled! macros, or any direct dispatch.enabled() that isn't followed by a dispatch. It does not fix the re-entrancy triggers (#2448, #2704), where a layer's on_event emits its own event and the inner pass's enabled() clobbers FilterState bits the outer pass hasn't consumed yet. That case has enabled() called fresh; the staleness is from nesting, not skipping, and fixing it needs FilterState to be a stack (or the API rework #3516 discusses). This PR doesn't attempt that — it makes the existing mechanism sound for the non-re-entrant case at the cost of the fast-path, and leaves the deeper fix for #3516.

Testing. New tests/layer_filters/stale_filter_state.rs reproduces #2519: two per-layer-filtered layers, a span both accept (cached always before this PR), then a bare dispatch.enabled() both reject, then a second span both accept. Before this PR the second span never reached either layer's on_new_span; with it, both spans do.