feat(orchestration): bound agent loops with caps + escalation (#9)

[andybaran]

Closes #9.

[andybaran]

🔬 Code review (code-review agent)

Verdict: APPROVED — 0 Critical, 0 Important, 0 Minor.

• Scope clean: only SKILL.md, references/agent-prompts.md, evals/evals.json.
• Numeric caps (max 3 rounds) for BOTH plan↔review and code-review↔implement, consistent across Step 2, Step 3.5, Quick reference, and agent-prompts.md. Old advisory "~3 rounds" replaced with a visible round N/3 counter + hard stop.
• Escalation concrete: ## ⛔ Escalation: … did not converge posted via gh issue comment heredoc, with summary + human hand-off; forbids silent loop AND silent give-up.
• Status-code short-circuit correct: BLOCKED/NEEDS_CONTEXT break immediately (consistent with Orchestration upgrade: blocking code-review agent role, implementer status codes, receiving-review guidance #6). No-progress guard framed as substantive comparison. Human-final-approval gate interaction sound.
• evals.json parses; scenarios 0–8 preserved; new scenario plan-review-nonconvergence-escalates asserts ≤3 review comments + escalation. Links resolve.

Human orchestrator performing final squash-merge.