docs(skills): modularize governance references - #670

Problem

Several workspace skills keep their full methodology, examples, failure histories, commands, and delivery rules in one large SKILL.md. Agents must load all of that material even when only one phase is relevant, while the long files make routing rules harder to review and maintain.

Solution

Turn each main SKILL.md into a concise router and move detailed guidance into named references/ files for architecture audit, dual-instance verification, E2E testing, performance guard, and React best practices. Update .gitignore so the newly vendored architecture and E2E references are tracked. During integration with current develop, preserve the newly strengthened dual-instance fleet invariants and add a mandatory determinism reference covering two-boot comparison, liveness, fault injection, unexplained deltas, and baseline attribution.

Potential risks

Moving methodology text can accidentally omit a rule or leave a broken relative link, which would make an agent skip required evidence. The change preserves the main routing contracts, validates every referenced file, and explicitly reconciles the latest dual-instance additions rather than replacing them with the older modularized text. No runtime code, dependencies, persisted data, or public API changes are involved. Rollback is a documentation-only revert.

Verification

• Checked all five main skill files for YAML frontmatter containing name and description — passed.
• Scanned every Markdown references/... link under .orgii/skills and verified its target exists — passed.
• git diff --cached --check — passed.
• Compared the staged file list with current origin/develop; only .gitignore and the five skill packages are included.
• Manually reconciled the dual-instance-verification conflict with current develop, retaining fleet-wide epoch/count invariants, destructive-effect auditing, resource anomaly handling, and mandatory determinism/fault-injection guidance.

No product tests or UI screenshots were run because this PR changes agent methodology Markdown and ignore rules only.