chore: release

[github-actions]

Package: dial9-trace-format-derive 0.4.1 -> 0.5.0

Package: dial9-trace-format 0.4.1 -> 0.5.0

Package: dial9-core 0.4.0-alpha.1 -> 0.4.0-alpha.1

Package: dial9-macro 0.3.7 -> 0.4.0

Package: dial9-utils 0.4.0-alpha.1 -> 0.4.0-alpha.1

0.4.0-alpha.2 - 2026-07-17

Package: dial9-tokio-telemetry 0.4.0-alpha.1 -> 0.4.0-alpha.2

Changes:

Added

• dial9::main now enables graceful_shutdown (#527)
• (config) [breaking] in-memory support for dial9 main macro (#490)
• [breaking] replace TelemetryEvent and Rust decode side with new serde-based approach (#485)

Fixed

• Avoid truncation when rendering the s3 browser view (#582)

Other

• move memory profiling to dial9-perf-self-profile (#591)
• move event bus to dial9-core (#549)
• setup dial9-core (#540)
• Drop aws-sdk-s3-transfer-manager, upload segments via aws-sdk-s3 PutObject (#668)
• (deps) bump s3s crates to 0.14.1 to pull in patched quick-xml (#611)
• [breaking] collapse handles into Dial9Handle / Dial9TokioHandle (#535)
• (config) Remove legacy positional config (#538)
• overlap trace download and parse (streaming decode + load-timing UX) (#568)
• Support conventional unit suffixes for viewer (#516)
• Cache the symbolizer across segments (#462) (#465)
• Add dial9-html-report skill: agent-authored HTML trace reports (#488)
• Improve Wire ID Resolution (#487)

Added

• Explicit CPU profiling backend selection via CpuProfilingConfig::with_perf_backend() and CpuProfilingConfig::with_ctimer_backend() constructors. The default (Auto: try perf, fall back to ctimer) is unchanged (#579, #660)
• In-memory writer (InMemoryWriter): run the trace pipeline with no filesystem dependency, encoded segments are held in process memory and shipped by the existing processor pipeline (#435)
• #[dial9_tokio_telemetry::main] now performs an implicit graceful shutdown after the async body returns: it drops the runtime and drains the background worker so the final segment is symbolized, compressed, and uploaded. Configure the deadline with Dial9Config::builder()...graceful_shutdown(Duration) (default 1s) or skip it with .disable_graceful_shutdown(). The low-level TracedRuntime API is unchanged — call TelemetryGuard::graceful_shutdown yourself (#479)
• SegmentProcessor::finalize_dump: custom processors can flush per-dump state when an on-demand dump completes (#549)
• ProcessError::into_parts: destructure a processing error into its kind and the carried segment data (#549)
• Memory profiling can run standalone, without dial9-core: depend on dial9-perf-self-profile with the memory-profiling feature, set SamplingAllocator as the global allocator, call MemoryProfiler::install(), and drain MemorySamples via MemorySampler::drain. The dial9 trace integration is that crate's dial9-source feature.

Changed

• Programs using #[dial9_tokio_telemetry::main] with the fluent Dial9Config now drain the telemetry worker on clean exit (up to the graceful-shutdown deadline, default 1s) instead of exiting immediately. This adds a bounded amount of shutdown latency but ensures the final segment is processed; opt out with .disable_graceful_shutdown(). The deprecated positional config is unaffected (#479)
• Breaking: renamed RotatingWriter to DiskWriter. The writer is now generic over its storage backend (SegmentWriter<Mode>) with DiskWriter / InMemoryWriter as the public types; memory constructors are InMemoryWriter::new / ::builder (#435)
• Breaking: SegmentData::segment() returns &SegmentRef (disk- or memory-backed) instead of &SealedSegment; processors that read the segment path must match the enum (#435)
• Breaking: collapsed the handle types into Dial9Handle (record/control, runtime-agnostic) and Dial9TokioHandle (spawn only). TelemetryHandle and RuntimeTelemetryHandle are removed (#535)
• Breaking: recording is now a method: record_event(event, &handle) becomes handle.record_event(event) (#535)
• Breaking: guard.handle() now returns Dial9Handle (record/control). To spawn instrumented tasks use Dial9TokioHandle::current(), guard.tokio_handle(&runtime), or the handle from trace_runtime().build() (#535)
• Breaking: boot_id is no longer an S3Config builder field. The runtime injects the on-disk namespace boot_id into the S3 config at build time, so a local trace segment and its S3 key share one identity. An S3Config built outside the managed Dial9Config path falls back to a fresh {4-alpha}-{pid} (#566)
• Memory profiling moved into dial9-perf-self-profile. The dial9_tokio_telemetry::memory_profiling::* paths are preserved by re-export, so most code is unaffected.
• Breaking: MemoryProfiler::install no longer takes a Dial9Handle, it returns a MemorySampler for standalone draining. For the dial9 trace integration use MemoryProfiler::install_into(&handle) (#591)
• Breaking: removed MemoryProfilerGuard. Memory profiling now drains for the session/guard lifetime (like CPU profiling) rather than being a permanent process-wide install (#591)
• Breaking: renamed Dial9Allocator to SamplingAllocator, the global-allocator wrapper now lives in dial9-perf-self-profile and works without dial9 (#591)

Fixed

• Viewer: browsing a nonexistent bucket now returns HTTP 404 instead of 500, and a syntactically invalid bucket name returns HTTP 400. The S3 NoSuchBucket/NoSuchKey and InvalidBucketName error codes were falling through to the generic error arm, which logged an "unclassified S3 error" and reported a server fault — so a user typo in the bucket name polluted the viewer's fault metric. They now classify as NotFound (404) and BadRequest (400) respectively.

Package: dial9 0.3.13 -> 0.3.14

---

Maintainer note: CI does not trigger automatically on this PR (GitHub limitation with GITHUB_TOKEN-created PRs). Close and reopen this PR to trigger CI before merging.