fix: address review findings in span processor API refactor

- Fix end_with_timestamp bug: use explicit end_time_set flag instead of
  sentinel comparison (start_time == end_time), which silently overwrote
  user-provided timestamps equal to start time
- Restore mem::replace for span_context in end_and_export to avoid
  unnecessary clone of TraceState (span is dead after data.take())
- Rewrite end_and_export to take data before shutdown check, preventing
  potential data leak on re-entrant Drop
- Add FinishedSpan::is_consumed() method and document post-consume behavior
- Add instrumentation_scope() to ReadableSpan trait (OnceLock for MSRV 1.75)
- Fix FinishedSpan::Debug to always show consumed=true when span data is gone
- Add ReadableSpan trait doc explaining post-end behavior on Span
- Use InMemorySpanExporter in test to verify end_time preservation
- Normalize formatting in ReadableSpan impl for FinishedSpan
- Remove redundant imports, fix doc typo, remove double blank line
- Update benchmark results for Apple M4 Max

Author: bryantbiggs

opentelemetry-sdk/CHANGELOG.md

@@ -28,11 +28,16 @@
 - **Breaking** `SpanProcessor::on_end` now takes `&mut FinishedSpan` instead of `SpanData`.
   Processors that only need to read span data can use the `ReadableSpan` trait without cloning.
   Processors that need ownership call `FinishedSpan::consume()` — the last processor receives
-  ownership via move (zero-copy), earlier processors receive a clone.
+  ownership via move (zero-copy), earlier processors receive a clone. Both `Span` (live) and
+  `FinishedSpan` (ended) implement `ReadableSpan`, enabling processors to inspect span data
+  in both `on_start` and `on_end` hooks. New `FinishedSpan::is_consumed()` method allows
+  checking whether span data has been consumed.
   Supersedes [#2962](https://github.com/open-telemetry/opentelemetry-rust/pull/2962).
   Relates to [#2940](https://github.com/open-telemetry/opentelemetry-rust/issues/2940),
   [#2726](https://github.com/open-telemetry/opentelemetry-rust/issues/2726),
   [#2939](https://github.com/open-telemetry/opentelemetry-rust/issues/2939).
+- Fix `Span::end_with_timestamp` preserving explicit end times even when equal to start time,
+  instead of silently overwriting with the current time.
 
 [3227]: https://github.com/open-telemetry/opentelemetry-rust/pull/3227
 [3277]: https://github.com/open-telemetry/opentelemetry-rust/pull/3277

opentelemetry-sdk/Cargo.toml

@@ -130,7 +130,6 @@ name = "span_processor_api"
 harness = false
 required-features = ["testing"]
 
-
 [lib]
 bench = false
 

opentelemetry-sdk/benches/span_processor_api.rs

@@ -12,17 +12,17 @@ use pprof::criterion::{Output, PProfProfiler};
 
 /*
 Adding results in comments for a quick reference.
-  Chip:	Apple M1 Max
-  Total Number of Cores:	10 (8 performance and 2 efficiency)
+  Chip:  Apple M4 Max
+  Total Number of Cores: 16 (12 performance and 4 efficiency)
 
 SpanProcessorApi/0_processors
-    time:   [385.15 ns 386.14 ns 387.25 ns]
+    time:   [190.51 ns 190.82 ns 191.13 ns]
 SpanProcessorApi/1_processors
-    time:   [385.73 ns 387.17 ns 388.85 ns]
+    time:   [191.60 ns 192.53 ns 193.42 ns]
 SpanProcessorApi/2_processors
-    time:   [384.84 ns 385.66 ns 386.50 ns]
+    time:   [191.24 ns 191.67 ns 192.15 ns]
 SpanProcessorApi/4_processors
-    time:   [386.78 ns 388.17 ns 389.58 ns]
+    time:   [192.94 ns 193.48 ns 194.12 ns]
 */
 
 #[derive(Debug)]