feat(openclaw): custom SpanProcessor injection + fix runaway ENTRY/AGENT span durations

[fangxiu-wf]

Closes #54

Summary

Two changes to opentelemetry-instrumentation-openclaw, released together as 0.1.5-beta.

Feature — custom SpanProcessor injection

• New config spanProcessorModule (env fallback ARMS_SPAN_PROCESSOR_MODULE): points at a module whose default export is a SpanProcessor. Absolute paths as-is; relative resolved against OPENCLAW_HOME (~/.openclaw).
• Package subpath export @loongsuite/opentelemetry-instrumentation-openclaw/span-processor provides defineGenAiSpanProcessor, dispatching by GenAI span kind (onLlmEnding/onToolEnding/onAgentEnding/onStepEnding/onEntryEnding) and hiding semconv dialect differences.
• Graceful degradation: load/validation failure falls back to the built-in processor; every user callback is try/catch-isolated so a faulty processor can never break the export pipeline.
• Fills the gap that globalSpanAttributes only supports static values.

Fix — runaway ENTRY/AGENT span durations (seen as 1d+ traces)

• Root cause: agent_end attributed context purely by channel. When its channel differs from the one llm_input registered (system/ vs agent/ mismatch), it failed to find the turn's context and closed nothing, leaking the open ENTRY/AGENT spans. A later agent_end (or session reset) then mis-closed them with Date.now(), producing day-long durations.
• Primary fix: when channel resolution yields no closable span, recover the turn's context via the runId openclaw attaches to agent_end (channel-first, runId as supplement — the common path is unchanged).
• Hardening: the stale-context sweeper now force-closes still-open ENTRY/AGENT/STEP spans using a bounded last-activity time (never Date.now()) and fully cleans up all lookup structures.
• Also fixes tool spans silently dropped on the same channel mismatch (before_tool_call recovers via runId, with a guard for the runId==sessionId closing-turn case).
• New diagnostic attribute on ENTRY: openclaw.trace.close_reason ∈ {normal, runid_recovered, stale_sweeper}.

Test plan

• tsc --noEmit clean
• Unit tests 143/143 pass — new: span-processor.test.ts (dispatch / dialect tolerance / graceful degradation / runtime isolation), channel-mismatch-repro.test.ts (tool-span recovery), agent-end-stale-close.test.ts (agent_end runId recovery + sweeper bounded force-close)
• Real ARMS link verified: healthy traces report close_reason=normal with sane second-level ENTRY/AGENT durations and intact hierarchy (no regression); a custom SpanProcessor was injected and confirmed to add business.* attributes onto LLM/ENTRY/TOOL spans in ARMS.

🤖 Generated with Claude Code