Release v0.8.0: MCP and skill fixes, compatibility checks and hooks

[JRS1986]

Release v0.8.0 fixes stale generated agent configuration and closes gaps in MCP discovery and skill approvals. It adds offline compatibility checks with dated provenance and optional Codex/Claude lifecycle hooks. The package version, lockfile and changelog are updated together.

Agentic coding disclosure

• Agent / tool used: Codex.
• Model / provider: GPT-6 / OpenAI.
• Human operator: @JRS1986.
• Task: Implement the requested MCP, skill, compatibility and hook improvements, then commit, push and release them.

What changed

• Files changed: See diff. MCP/eval validation, native skill discovery and complete-package approvals, compatibility provenance/checks, lifecycle hooks, CLI/doctor integration, adapter and model catalog refresh, regression fixtures, docs, and v0.8.0 metadata. Generated local graphify caches are ignored.
• Commands run: uv lock; uv sync --locked --extra dev; uv run --offline --locked ruff check; uv run --offline --locked pytest -q; uv run --offline --locked coding-scaffold context lint --target .; uv run --offline --locked coding-scaffold eval run --target .; uv run --offline --locked coding-scaffold tools compatibility --target .; uv build --offline; npm run build in docs/; isolated built-wheel installation and CLI smoke checks; git diff --check. UV commands used a writable temporary cache.
• Tests run: 734 tests passed; Ruff and context lint passed; readiness passed 9/9. Compatibility found zero errors and the expected warning that this generator repository has no saved compatibility manifest. Wheel/source distribution and docs builds succeeded. Installed-wheel smoke checks verified version 0.8.0, module imports, hook generation, skill creation and compatibility reporting.
• Tests not run and why: Live Codex/Claude hook execution was not exercised; hook contracts, payloads, preservation, privacy and loop guards are covered offline. GitHub's Python 3.11–3.14 matrix will run before merge.

Risk surface

• External tools or MCP servers used: Shell, uv, npm, Git and GitHub CLI; web research of official upstream documentation during implementation. No MCP server is added to this repository.
• Secrets / data exposure risk: No credentials or customer data added to the change. Hook reports contain aggregate counts and a timestamp, never transcripts. Existing graphify output with local paths stays local.
• Network calls made during the session: Upstream documentation research, package metadata/build dependencies, and GitHub fetch/push/PR/release operations. No runtime network, model calls, telemetry or new runtime dependencies.

Review focus

• Human review focus: Malformed MCP configs fail readiness; approvals cover every package file and executable bit; hooks merge without replacing existing handlers and limit Stop continuation to one attempt.
• Known limitations: Compatibility validates reviewed fields offline rather than installed binaries or live model availability. Metadata parsing supports a documented YAML scalar subset. Checksums are unsigned drift markers. Hooks remain opt-in and do not execute project tests. Legacy flat aliases remain available with deprecation warnings until 0.9.0. The existing CLI error unification changes the Python exception type to CliError, as documented in the changelog. The local docs build emitted a non-fatal Rspack cache-write warning.
• Reusable knowledge captured: Session trace, Tool Adapters, Upgrading, and Security.