chore(workflow): replace Codex agent layout with Claude Code

[kaikisegfault]

Outcome and scope

The repository targeted Codex sessions. This moves the agent contract, skills, and session configuration to their Claude Code equivalents so a clean session opened at the repository root loads the correct instructions. No issue: this is a single coherent tooling slice with no acceptance criteria to negotiate.

Before	After
AGENTS.md	CLAUDE.md
.agents/skills/*/SKILL.md	.claude/skills/*/SKILL.md
.agents/skills/*/agents/openai.yaml	removed (Codex-only interface metadata)
.codex/config.toml	.claude/settings.json

Tooling follows the layout rather than only its names:

• tools/verify_metadata.py drops interface_metadata() and tomllib, validates .claude/settings.json as JSON, and no longer requires agents/openai.yaml per skill.
• tools/verification_scope.py drops the .agents/skills/**/agents/openai.yaml lightweight-path rule.
• .github/CODEOWNERS, README.md, and five documents under docs/ follow the rename.
• .gitignore now excludes /.claude/settings.local.json.

Protocol, compatibility, economic, and security effects

None. No consensus-visible type, encoding, state transition, economic rule, or cryptographic behavior is touched. The change is confined to agent-session tooling, repository documentation, and their validators.

Verification

$ python3 tools/verify_metadata.py
Validated 5 repository skills and 19 internal Markdown links.

$ python3 tests/tools/verify_metadata_test.py
Ran 3 tests — OK

$ python3 tests/tools/verification_scope_test.py
Ran 3 tests — OK

Scope classification behavior confirmed unchanged in intent:

README.md + .claude/skills/proceed-project/SKILL.md  -> lightweight
README.md + .claude/settings.json                    -> full

.claude/settings.json remains a control file that fails closed to the full hosted matrix, exactly as .codex/config.toml did; SKILL.md edits remain on the focused metadata path.

Known limitations and follow-up

• verify_metadata_test.py now asserts >= 2 errors instead of >= 3 in the rejection case, because removing the openai.yaml default_prompt check removes one expected error. The remaining two (name mismatch, TODO marker) are the assertions that carry review value.
• Skill instruction bodies were renamed mechanically from "Codex" to "Claude Code". Their content is otherwise unchanged and may deserve a separate pass for Claude-specific phrasing.