Unavailable observations become false changes in baseline comparisons #100

Problem and proposed fix

Unavailable or missing collector sections could be interpreted as measured zeros or empty sets, creating false connection drops, standby losses, new-query entries and archiver events. Require observable section provenance at both ends while preserving genuinely measured empty results and positive evidence in legacy snapshots. A retained positive connection gauge remains usable when only rates are unavailable.

Verification

• Regressions fail on unchanged base behavior for missing current observations, missing baselines and empty legacy sections.
• The real SQLite Save/reopen/offline-diff path reproduces false changes and passes with the fix.
• Affected diff/CLI race suites, clean committed-HEAD bash scripts/gate.sh, and go test -race ./... pass.
• Controls preserve genuine losses/additions, measured zero, legacy positive values, rate-reset connection gauges, deterministic output and input immutability.

Scope

No schema, SQL, dependency or migration changes. This suppresses comparisons unsupported by observations, not genuine measurements. It is distinct from #64 (cross-reset comparability), #69 (standby multiplicity) and #87 (stored scalar extraction), though the diff edits may need reconciliation as those land. No live PostgreSQL failure is claimed. Revert the commit to restore the former comparison rules.

I would like to address this with the focused fix and regression coverage described above. Existing issues and open PRs were checked for overlap before preparing the change.