Bandingkan sumber

Source conflict detection (Sprint 5). Two series that claim to measure the same thing are compared;
they are never merged, and neither is declared right.
Candidate pairs:
    declared      series people listed as equivalent (config/data_lineage.yaml `equivalents`): the conflict list
    same id       two stored series with the same provider and provider series id (catalog): listed separately
                  for review, because two keys can read different fields of one provider table (a daily flow
                  and its running total share an id), so a shared id alone does not make them the same thing
Any two series can also be compared on request. The cockpit file's own cross-checks and the collectors'
parity check are separate pages (Research > Tools) and are linked, not recomputed.
For a pair: frequency, unit, history and, on the dates both have, B minus A. Values "differ" only beyond
rounding: half a unit in the last decimal place of the coarser series (from the stored values, up to 6
places). If more values agree after moving B by a few days (up to SHIFT_DAYS either way), the pair is
reported as using different date conventions.
A FORECAST or CONSENSUS series and an ACTUAL one are not competing observations: such a pair is shown
with that label and never listed as a conflict.