What happens to a span with no call-site id on it?

It still ingests fully. Tessary doesn’t require tessary.call_site.id to accept a span, and resolution is explicit or nothing: the tag or no call site, never a guess, so an untagged span resolves to no call site rather than a guessed one. What changes is what you lose: call-site-scoped features, drift baselines, per-call-site assessment, and the evidence triage and RCA reason over, skip it, because they need a place to accrue history to. In a multi-step chain, a step nobody tagged is effectively invisible to that binding, even though the data behind it is there and stored; a regression born in that step surfaces attached to whatever the untagged spans carry, not the step that actually caused it, which makes triage slower, not impossible. Tagging is the fix, and it only applies going forward from whenever you add it. Nothing you didn’t tag gets backfilled.