feat: versioned deploys with update subcommand + MCP server plan

[bebingando]

Summary

• Introduces a versioned deploy model: every release carries a semver (see root VERSION, seeded at 0.1.0), and deploy.sh / deploy.ps1 write a .agentic-context.lock at the target recording installed version, agent set, and per-file SHA-256 + ownership.
• Adds an update subcommand (./deploy.sh update <target> / .\deploy.ps1 -Update) that refreshes template files whose local hash still matches the lockfile, while preserving locally-edited files and never touching project-owned files (AGENTS.md, CLAUDE.md, .claude/settings.json).
• Documents the roadmap for an optional MCP server in mcp-server/README.md — TypeScript + @modelcontextprotocol/sdk, same-repo subdirectory reading from the existing standards/ and playbooks/ trees, distributed via npx. Additive, not a replacement; copy-deploy remains the universal fallback.

Motivation

The current copy-deploy is a defensible shadcn-style model, but the recent overwrite-guard commit was the first symptom of its ceiling: once files are copied, there is no way to distinguish a pristine template from a user-customised one, and no way to propagate upstream improvements without clobbering. This PR adds that distinction as the smallest-possible primitive — a lockfile + an ownership classification — so re-running deploy.sh in update mode becomes a safe merge rather than a conflict-prone re-init.

MCP is the natural next step for "library of context many agents read", but it's not universally supported across agents yet. The plan is sequenced deliberately: versioning + update now (no runtime dependency, works with every agent), MCP server later (opt-in, additive, same content). The mcp-server/ README captures the architecture so a contributor can execute that phase without re-deriving the design.

Ownership model

Ownership	Meaning	Examples
template	The upstream repo owns the content. Auto-refreshed on update if the target copy is pristine.	.context/standards/*, .context/playbooks/*, .windsurfrules, generated .claude/skills/*
configure	The target repo owns the content (project-specific [CONFIGURE] sections or permissions). Written once on init; never overwritten on update.	AGENTS.md, CLAUDE.md, .claude/settings.json

Behaviour

• Clean update (no changes anywhere) — silent, reports "No template files needed updating."
• Upstream changed, local pristine — refreshed silently, listed under Updated.
• User edited a template file — skipped, listed under Preserved — locally modified template files, with guidance on how to accept upstream or keep local edits.
• New file added upstream — copied in, listed under Added.
• Configure file — always reported under Preserved — project-owned, never touched.
• --overwrite / --no-overwrite are rejected in update mode (they don't apply).

Test plan

Tested end-to-end locally against throwaway target directories. PowerShell parity implemented but not runtime-verified (no pwsh on this machine) — mirrors the bash logic closely and should be smoke-tested on Windows before merge.

• Fresh ./deploy.sh --agents claude <target> writes lockfile with 109 tracked files; ownership classification correct (AGENTS.md, CLAUDE.md, .claude/settings.json → configure; everything else → template).
• ./deploy.sh --agents all <target> writes lockfile with 143 tracked files.
• Subsequent update on unmodified deployment reports no changes.
• Edit .context/standards/docker.md locally, run update → file preserved with warning, lockfile entry unchanged.
• Edit a source standard in the agentic-context repo, run update → pristine local copy refreshed, listed under Updated.
• Add a new playbook to the source, run update → both the playbook file and its generated skill wrapper appear under Added.
• Edit AGENTS.md locally, run update → untouched, listed under Preserved configure.
• PowerShell: verify .\deploy.ps1 -Update -TargetRepo <path> on Windows.

Files

• VERSION — new, seeded at 0.1.0.
• deploy.sh — adds lockfile infra (read/write_lockfile, compute_hash, ownership_for), dispatches copy_file and generate_skill on DEPLOY_MODE, adds update subcommand parsing and update-mode main flow + summary.
• deploy.ps1 — same treatment: Set-LockEntry / Get-LockHash / Read-ExistingLockfile / Write-Lockfile, -Update switch, Copy-SingleFile and New-SkillWrapper dispatch on $script:DeployMode.
• README.md — adds Versioning & Updates section covering the ownership model, the lockfile, and the workflow.
• mcp-server/README.md — new; captures the design for the "later" phase.

No changes to standards, playbooks, conventions, AGENTS.md, or the agent redirect stubs.

🤖 Generated with Claude Code