SDR-0012: Protect missing event windows instead of silently reconstructing them

• Status: Accepted
• Date: 2026-07-26

Context

Published photometry workflows variously interpolate isolated dropped frames, replace selected artifact regions, remove and concatenate segments, or downsample without reporting consequences for event summaries. The frozen v0.1.1 benchmark found that reconstruction safety depends on transient shape and gap location: even one missing event-adjacent sample and two-sample gaps caused retained failures.

Decision

Linear median-rate regularization is an experimental, declared policy for small timestamp jitter. Missing fluorescence samples are different: contiguous gaps are never silently bridged, and isolated samples inside inferential windows default to protected missingness. Linear isolated-sample reconstruction may appear only as an explicit sensitivity universe.

Event summaries require complete finite baseline and response windows. They report finite coverage, reconstructed coverage, and an event disposition. Peri-event alignment interpolates only within finite contiguous runs. Condition-dependent exclusion or reconstruction produces a warning.

Consequences

Some datasets will retain fewer analyzable events, but no partial event is treated as fully observed. Display interpolation must be separately tagged in any future implementation. The original values and masks remain available for alternative prospective policies.

Revisit trigger

Revisit after validation against real held-out dropouts, sensor-specific transient libraries, or a reconstruction model that supplies calibrated uncertainty rather than a single filled trace.