docs: strengthen agent-instruction guardrails (session-start gate, change-scope rules, thin AGENTS/CLAUDE pointers)

[ernysans]

Summary

Documentation-only improvement pass on this repo's agent-instruction files, bringing them up to a guardrail bar established across sibling repos. No application/library code or build output (lib/) changed.

Changes

• Session Start Identity Gate (.github/copilot-instructions.md §0) — new: at the start of every workflow, before acting, ask whether the requester is the project owner and whether they have permission for destructive changes; non-owners are restricted to small, narrowly-scoped fixes and directed to open a Task/Issue for larger work; only a top-tier agent may work with a non-owner.
• Change Scope Guardrails (§0.1) and Deployment out-of-scope note (§0.2) — surgical-changes-only policy, explicit reminder that a src/ change isn't complete until npm run build output is regenerated and committed together (per this repo's existing CI drift check), and a plain statement that this repo has no publish/deploy automation (verified against .github/workflows/nodejs.yml).
• De-named all specific AI model references across instruction files in favor of tier-neutral wording (routing/escalation policy preserved); removed human-effort/hour estimate tracking.
• Added root AGENTS.md and CLAUDE.md as thin pointers to the canonical .github/copilot-instructions.md and .github/instructions/*.md files (no restated rule text, so nothing can drift out of sync).
• One-line cross-reference pointer added in .github/instructions/cross-repo.instructions.md §6.

Verification

• gh api repos/furcata/core-node --jq '.visibility' → public (confirms nothing furcata-internal may be referenced).
• Repo-wide grep for named models / human-hour estimates / private furcata repo references → zero hits.
• src/ and lib/ untouched — no rebuild required, build parity intact.