Propagate worker user span context

[mar-cf]

No description provided.

[ask-bonk]

After thorough review, I've checked:

1. Autogate: Correctly added with enum + string mapping + guarding usage. Follows conventions.
2. SubrequestMetadata: New userSpanContext field added, all call sites updated consistently.
3. getUserSpanContext: Autogate-gated, kj::downcast<UserSpanObserver> is safe (only one SpanObserver subclass in codebase, guarded by KJ_IF_SOME so only entered when observer exists).
4. SpanContext::clone refactoring: Changed from member to static — no remaining callers of the old form. TailEvent::clone() constructs SpanContext directly, doesn't use SpanContext::clone().
5. TraceContext::getUserSpanParent(): New accessor, consistent with existing getInternalSpanParent().
6. Onset parentSpanId: Now correctly populated from trigger context's spanId.
7. server.c++ triggerContext construction: Correctly handles Maybe<SpanId> from SpanContext::getSpanId(), constructs trigger InvocationSpanContext which propagates traceId to child.
8. UserSpanObserver::getSpanId(): Simple accessor for existing field, fine.
9. Test rewrite: Well-structured tree-based comparison replacing flat array. Correctly handles both autogate on/off variants. The buildTree logic properly scopes by traceId to avoid false matches.
10. Memory/thread safety: No new ownership issues. SpanContext is value type, safely passed through SubrequestMetadata. No cross-thread concerns.
11. Backward compat: Gated behind autogate, so no behavior change without explicit opt-in.

Everything looks clean and well-structured. No actionable issues found.

LGTM

github run

[jasnell]

this will need to be rebased on main for CI to pass.

[mar-cf]

Confirmed with https://github.com/danlapid/stw-local-dev

Will post when I exercise all the different paths

[ask-bonk]

I've completed a thorough review of all 15 changed files. Let me summarize my findings:

Autogate: Correctly registered with enum + string mapping in sync. All behavioral changes are gated behind USER_SPAN_CONTEXT_PROPAGATION.

Backward compatibility: Fully gated behind autogate. When disabled, traceId is nullptr, so toSpanContext() returns kj::none, and no trigger context is created - preserving existing behavior exactly.

Memory/thread safety:

• TraceId is trivially copyable (defaulted copy ctor), so lvalue copies in toSpanContext() and newChild() are safe and don't zero the source field.
• SpanContext is move-only but the new static clone() correctly constructs a new one from const ref.
• No cross-thread concerns - all changes stay within the same request flow.

API design: The SpanParent::toSpanContext() virtual dispatch through SpanObserver is clean. The interface mirrors the existing getInternalSpanParent() / getUserSpanParent() pattern.

Test: Well-structured tree-based comparison. buildTree() correctly scopes by traceId to avoid false matches. Both autogate-on and autogate-off paths are tested via the @ and @all-autogates variants. The unsafe_module flag addition is needed for workerd:unsafe import.

danlapid's review comment: Was on an older revision; the current code already uses .userSpanParent = tracing.getUserSpanParent() as suggested.

No actionable issues found.

LGTM

github run

[mar-cf]

Using only public repos, this and https://github.com/danlapid/stw-local-dev: