feat(workflow): experimental self-evolving workflow loop

[justrach]

What

Adds the workflow orchestrator agent + a workflow-evolver meta-agent that closes a run -> telemetry -> learnings loop, plus helper scripts. Experimental.

• workflow (built-in) — decomposes a goal, fans out parallel task sub-agents (sage read-only, forge for edits), synthesizes. Carries a machine-owned Learnings section + emits a run-telemetry block.
• workflow-evolver (built-in) — reads accumulated run telemetry and proposes evolved Learnings; a deterministic splice (_evolve_apply.py) applies them between markers (version bump + changelog), with a constitution guard (refuses self-sabotaging learnings: disabling telemetry, suppressing task, hiding failures) and an append-only generation archive.
• scripts/ — run-workflow.sh (capture telemetry), evolve-workflow.sh + _evolve_apply.py (read-only evolver -> deterministic splice).

Honest caveats (why experimental)

• Telemetry is currently self-reported by the agent — gameable, per the Darwin-Goedel-Machine analysis. The sound version sources the same objective signal from the data.codegraff.com trajectory feed every workflow run already flows to (tagged agent_id); that rewire is the obvious follow-up.
• The loop evolves a project-local .forge/agents/workflow.md drop-in (built-in = frozen baseline, drop-in = evolving instance). Runtime artifacts (telemetry jsonl, evolution log, generations) are gitignored.
• No selection/regression gate yet (mutate-and-accept).

Verification

cargo check -p forge_repo clean (all 5 agents embed). The workflow orchestrator is verified end-to-end on a built binary (fans out sage/forge sub-agents, no hangs); the evolver loop was demonstrated v1->v3 on seeded + real telemetry.

🤖 Generated with Claude Code

[github-actions]

Action required: PR inactive for 5 days.
Status update or closure in 10 days.