Description
The k8s corpus producer records classic-histogram identity while eliding every le value, but the k8s-addons corpus retains real le values for equivalent families. Resolve the producer policy from observed egress so bucket evidence is represented consistently and downstream emitter work never chooses bounds.
Acceptance Criteria
- #1 The k8s corpus producer records observed le values and bucket bounds for classic histograms, or a documented evidence rule explains and consistently represents why a value cannot be retained
- #2 Instrument type and classic-histogram classification come from observed series structure, never a metric-name suffix
- #3 A regression fixture covers a real finite bucket and +Inf without exposing deployment identity
- #4 The refreshed corpus and comparator report make retained versus absent bucket evidence explicit
Definition of Done
- #1 just check (fmt-check, lint, gen-check, env-check, docs-check, test, race, hygiene, ui-check, compose-check, helm-test, lab-check, signal-fidelity)
- #2 just gen (only if a blueprint field, construct/workload config struct, or a skill under plugins/synthkit/skills/ changed)
- #3 just dump — inventory diffed against signals/
Implementation Plan
2026-08-31 Lane F phase 1: establish observed-series histogram classification and explicit retained/absent le evidence with regression fixtures; request the root-scheduled exclusive lab capture before corpus refresh.
Implementation Notes
2026-08-31 exclusive alloy-default capture completed in 441s with a fixed 300s window: 107 metric families and 22,219 requests. storage_operation_duration_seconds had real mount/unmount activity but only the literal _count family; no _bucket, _sum, le, or histogram object reached egress. This satisfies the evidence dependency without choosing bounds.
Campaign label 2026-08-31; the exclusive capture actually executed on 2026-08-30. A fixed 300-second window produced 22,219 RW1 series. Observed-series structure now owns histogram classification, fixtures retain finite and +Inf evidence without identity, and absent bucket evidence remains explicit.
Final Summary
2026-08-31: Corrected histogram evidence handling to use observed series structure and refreshed the k3d corpus from a five-minute capture. The storage family had real activity but no bucket, sum, histogram block, or le evidence; no bounds were chosen.