Fix duplicate failed-step log attribution

[cursor]

Description

Fixes pipeline diagnosis for matrix, loop, and repeated steps that share the same stage/step identifier. Failed logs now retain the execution node ID, colliding display keys are preserved instead of overwritten, and all_step_logs merges already-fetched content by the unique logBaseKey.

Without this fix, one failed node's log could overwrite another and then be attached to both node-ID-keyed entries, causing pipeline summaries to report incorrect failure evidence.

Adds a regression with duplicate failed step identifiers and distinct log content.

Type of Change

• Bug fix
• New feature
• Refactor
• Documentation
• Other

Checklist

• pnpm test passes
• pnpm typecheck passes
• pnpm build passes
• pnpm standards:check passes (architecture guardrails — see docs/coding-standards.md)
• pnpm docs:check passes (if registry/tool counts changed)

Coding Standards (registry-driven MCP model)

If this PR adds or changes Harness API coverage:

• No new server.registerTool() calls — only toolset definitions in src/registry/toolsets/
• Toolset registered in ALL_TOOLSETS and ToolsetName union
• operationPolicy on every new/changed endpoint
• Shared response extractors from src/registry/extractors.ts (no raw passthrough on real endpoints)
• identifierFields and scope declared on new resources
• No console.log() in src/ (stdio JSON-RPC safety)

The API-coverage checklist is not applicable; this change is confined to diagnosis response assembly.

  

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}