Source quality and review

Source diagnosis and canonical processing have different responsibilities. Diagnosis reads source evidence and records findings. Processing persists cleaning actions and supported geometry in the same import transaction. Review reuses evidence for the matching source revision.

Reassessment does not reimport or run a second cleaning algorithm. Missing or mismatched evidence remains unavailable. Raw observation inspection keeps original source order and timestamps.

User inclusion decisions control map and statistical use. They cannot turn estimated timing into measured duration, speed, energy, or a time-based Health/media match.

Review thresholds

GpsReviewThresholds owns manual-review thresholds. REVIEW creates a pending source issue; warnings and routine findings remain informational. Assess each original source section, counting affected observations once. Coverage requires both the stated percentage and minimum count.

Retained abrupt position discontinuities keep their contextual detector: at least 200 m and 50 m/s, exceeding surrounding supported speed by more than six times. Edges require at least 125 m/s and stronger available-side support. An unresolved detected discontinuity requires review. Missing optional fields and fragment count alone do not create a manual review task. Generated timing and inconsistent processing results retain their safeguards.

A backward-clock alert alone does not disable a section's timing. Whole-section containment requires at least three independent backward transitions greater than one second covering at least 1% of source observations, a backward lag of at least 12 hours, or an unresolved large date excursion. Affected observations below an earlier clock maximum are not counted as independent transitions.

An isolated forward clock spike requires review as one incident. Import excludes only its derived timestamp when two neighbouring supported intervals establish a consistent return: each interval is positive and at most 120 seconds; the return spans at most 240 seconds and three times their maximum cadence; the spike exceeds both neighbours by more than one second and ten times that cadence. Missing context or smaller ambiguous peaks remain source evidence. Original timestamps remain in RAW. Persisted containment evidence keeps timing unsupported after reassessment until a successful reimport rebuilds it. Unsupported local intervals still make aggregate motion partial; retaining other source timestamps does not invent complete motion totals.

Aggregate changes in distance, ascent or descent after processing do not create review findings. Routine denoising can substantially change those totals. Individual correction magnitudes, rejection coverage and retained discontinuities remain assessed. Reassessment ignores stored aggregate-impact summaries, resolving their stored issues even when assessment is incomplete. Other unchecked findings remain active. It classifies current cleaning actions against source observations; it does not rerun the cleaner. New imports record the thresholds and processing version. Existing motion totals and missing cleaning evidence require ordinary reimport to rebuild.

Equal timestamps retain geometry within the same 200 m support bound as untimed observations. Their instantaneous speed remains unknown. The motion analyzer retains supported totals and uses a temporary monotonic timeline for clock precision reversals of at most one second, without modifying stored observations.

Diagnostics is a named module; the repair module has been removed. The Spring Boot application remains the assembly. General multi-module extraction is a separate proposal.

User procedures are in Source review; launch requirements are in GPS quality modules.

Assessment limits and clock alerts

Assessment has limits of 64 MiB and one million observations. The streaming parser retains observations and supported fields without building a document tree for device extensions. Nesting, field length, and retained review data have additional limits; unused extension elements do not consume the retained-data budget. An incomplete assessment is reported without treating the file as healthy or automatically excluding it. Original downloads are streamed without this limit; converted GPX downloads are bounded by it. All diagnosis runs on the server. Import also checks each source section for conflicting clocks when full assessment cannot finish.

A backward-time alert keeps the section's supported timing. Systemic clock problems disable timing for the affected section while retaining original observations, timestamps, and geometry. A clearly isolated forward clock spike excludes only that point's derived timestamp and appears as one incident; original timestamps remain available in RAW. Recognized timing problems remain visible after reassessment.

Abrupt retained relocations are checked against the surrounding pace. Endpoints use a stricter check with support from the available side. Unsupported connections become separate spans before smoothing; their connecting distance and speed do not contribute to statistics. Sustained fast travel alone does not trigger this check. Changes in total distance, ascent, or descent after smoothing do not trigger review; large individual corrections and substantial point rejection do.

Updating the server does not reassess existing findings. Reassess current file applies current criteria, removes earlier aggregate-impact findings, and reuses stored cleaning evidence only when its source revision matches. It preserves existing timing containment; an ordinary reimport is required to rebuild timing, motion totals, and missing processing evidence. Without import evidence, cleaning remains Not assessed.

Review does not edit, replace, or back up source files. Backup files from earlier versions remain on disk and can be managed outside the app.