feat: add JFR and performance engineering plugins

[jbachorik]

What & why

Adds two plugins to the marketplace for performance investigations that go beyond live BTrace probes:

• jfr-analyzer: turns JFR, pprof, OTLP and HPROF recordings into a structured investigation using USE (Utilization, Saturation, Errors) and TSA (Thread State Analysis) reasoning before looking at individual hotspots. It includes:
  • the jfr-analyzer skill with triage, drilldown, report and eval subskills;
  • the async-profiler-interop and jfr-btrace-interop skills;
  • a perf-engineer agent;
  • a Jafar MCP server entry (jfr-mcp, SSE on localhost:3000);
  • an eval corpus of 16 Java scenario apps with scoring scripts.
• perf-engineer: a session-driven performance engineering skill with evidence and gate references, plus a session script with unit tests.

It also registers both plugins in the Claude Code and Codex marketplaces and the Pi package, adds the design doc docs/perf-engineer-design.md, and links the BTrace observability skills to the new JFR interop workflow.

A second commit removes the allowed-tools lines these skills added. main's repository validator accepts only name and description in skill front matter, and the lines' unprefixed MCP tool names matched no tool in Claude Code anyway.

This is the base of a two-PR stack. The BTrace MCP fixes and e2e CI follow in a separate PR on top of this branch.

How verified

Gate	Result
scripts/validate-marketplace.sh on this branch merged into origin/main (includes validate-repository.js)	Repository consistency valid: 4 plugins, 22 skills, Marketplace validation passed.
node scripts/validate-evals.js	Eval corpus valid: 7 cases
python3 -m unittest test_perf_engineer_session (plugins/perf-engineer/scripts)	Ran 4 tests … OK
merge into origin/main	clean

Coverage notes

The validators check manifests, skill front matter and the BTrace observability eval corpus. The unit tests cover the perf-engineer session script. Nothing here exercises the skills' analysis quality.

Not verified / known limitations

• jfr-analyzer eval corpus not run. It needs async-profiler, a running Jafar MCP server and model calls, so it is meant to be run manually.
• jfr-analyzer MCP server not exercised. The plugin expects an externally started Jafar server at http://localhost:3000/mcp/sse.

Breaking changes / migration

None.

Links

Stack: base PR; the BTrace MCP fixes PR depends on this one.

🤖 Generated with Claude Code