PrecomputedSum in Delta temporality resets baseline on missed/rejected observations, double-counting values #3714

What happened?

In PrecomputedSum::delta, collection drains and resets the value_map. Any attribute set that is not observed in the current cycle (e.g. an async callback temporarily misses an observation, errors, or an invalid measurement like NaN is dropped) is omitted from new_reported.
When *reported = new_reported executes, all unobserved baselines are cleared.
When the instrument subsequently observes a new measurement, reported.get(...) returns None and defaults to 0. This causes the delta to be computed against 0 rather than the last valid baseline, re-counting previously exported values.

Reproduction:

1. Cycle 1: Observe 100 -> Collect: exports delta 100 (baseline cached as 100).
2. Cycle 2: Missed observation (or dropped NaN) -> Collect: exports nothing, but clears reported.
3. Cycle 3: Observe 110 -> Collect: exports delta 110 (110 - 0) instead of 10 (110 - 100), double-counting the initial 100.

What did you expect to happen?

Missed or rejected observations should not immediately discard the baseline. The next valid observation should compute the delta relative to the last reported value.

OpenTelemetry API Version (i.e version of opentelemetry crate)

0.32.0 (and main)

OpenTelemetry SDK Version (i.e version of opentelemetry_sdk crate)

0.32.1 (and main)

What Exporter(s) are you seeing the problem on?

No response

Relevant log output

Tip

_{React with 👍 to help prioritize this issue. Please use comments to provide useful context, avoiding +1 or me too, to help us triage it. Learn more here.}