chore: port cortexdj#48 — agent-config consolidation, Codex compat, lint/guard hooks

[LukeMainwaring]

Summary

Ports the agent-config + tooling improvements from cortexdj#48 to SampleSpace, adapted for terminology, the test layout, and frontend architecture differences. Four commits:

• d2f898b — pre-commit bump: ruff v0.14.10→v0.15.13, mypy v1.19.1→v2.1.0. The pinned hook versions were below the backend/pyproject.toml floors (ruff>=0.15.12, mypy>=2.1.0); this realigns them. pre-commit run --all-files passes — the mypy v1→v2 major bump introduced no new strict-mode errors.
• 17afbbb — AGENTS.md is now the tool-neutral canonical agent-instructions file (Claude Code, Codex, etc.). CLAUDE.md → symlink → AGENTS.md; .agents/skills → symlink → ../.claude/skills so Codex picks up the same skills. The @README.md/@DEVELOPMENT.md import syntax was dropped for plain prose (Codex cannot expand it).
• 8b1abba — new Claude Code hooks: backend-ruff.py + frontend-lint.py (PostToolUse), block-generated-edits.py (PreToolUse guard for frontend/api/generated/, fails closed when it can't determine a target path), wired in .claude/settings.json. Claude-Code-only by design — git pre-commit still covers backend lint on commit; these fill the inline + frontend + generated-guard gap for agent sessions.
• 4f7a881 — prompt/skill polish: code-reviewer.md (general first-pass framing, .claude/rules/ as single source of truth, new Tests dimension), vercel-chatbot-template.md (live tree discovery, version-compat check, scope constraints, re-adapted to SampleSpace's architecture), updating-deps/SKILL.md (concrete diff before commit, skip empty-diff).

Changes

• .pre-commit-config.yaml — ruff/mypy version bump
• AGENTS.md (new, canonical) / CLAUDE.md (→ symlink) / .agents/skills (→ symlink)
• .claude/hooks/{backend-ruff,frontend-lint,block-generated-edits}.py (new) + .claude/settings.json (hooks wired)
• .claude/agents/code-reviewer.md — deferred lint nits to hooks/pre-commit; dropped cortexdj's modal.md pointer (no Modal here); Tests dimension references the eval-marked suite in backend/tests/evals/
• .claude/agents/vercel-chatbot-template.md — re-adapted to SampleSpace's data-<name> DataPartRenderer architecture (not cortexdj's tool-<name> switch)
• .claude/skills/updating-deps/SKILL.md — commit section now shows a concrete diff and skips empty-diff runs

Caveats

• Intended behavior change: dropping the @README.md/@DEVELOPMENT.md imports means Claude Code no longer auto-injects those files; the agent reads them on demand. This is the cost of AGENTS.md tool-neutrality (same call cortexdj made).
• Symlinks are git-tracked as mode 120000 and degrade to plain text on a Windows checkout without core.symlinks. Accepted — macOS target environment.

Test Plan

• uv run --directory backend pre-commit run --all-files passes with bumped ruff + mypy
• CLAUDE.md resolves to AGENTS.md; .agents/skills resolves to .claude/skills (both tracked as mode 120000)
• All 3 hooks compile; block-generated-edits.py verified across 6 payload shapes (generated → block, normal → allow, no-path → fail-closed block, unparseable → allow)
• .claude/settings.json is valid JSON
• Sanity-check in a fresh Claude Code session that the PostToolUse hooks fire on a backend/frontend edit and the generated-edits guard blocks
• Spot-check Codex picks up AGENTS.md + .agents/skills

🤖 Generated with Claude Code