phase2-onboarding: adopt Phase 0 contract for m-stdlib-vscode

[rafael5]

Summary

Tier-2 onboarding per AI-discoverability-plan.md §3.4 and phase1-plan.md §9. This repo had no CLAUDE.md to rename — same starting state as m-test-engine PR #2 — so the AGENTS.md is authored from scratch.

What landed

Surface	Contents
AGENTS.md	Authored from scratch. Frontmatter declares the extension's marketplace id, manifest-discovery contract, the m-stdlib upstream dependency, and the sibling-extension relationship with tree-sitter-m-vscode. Five Phase-0 sections present (Setup / Test / Build / Verify / Guardrails). CLAUDE.md symlinked to it.
dist/extension-info.json	Hand-authored mirror of package.json — name, publisher, version (0.2.0), license, marketplace_id (rafael5.m-stdlib-vscode), engines (vscode ^1.85.0), the four m-stdlib.* settings keys, the four-step manifest_discovery_order, snippets_path. Sorted keys, 2-space indent, trailing newline.
dist/repo.meta.json	Phase-0 contract entry. Exposes extension_info (dist/extension-info.json) + package_json (package.json). consumes: ["tool:m-stdlib"]. License MIT. verification_commands: ["make check-manifest"].
tools/check-manifest.py	Stdlib-only validator — verbatim copy of the one in tree-sitter-m + m-test-engine.
Makefile	New. Three targets: manifest (informational — dist files are hand-authored, not regenerated), check-manifest, check-docs-prose.
.github/workflows/ci.yml	This repo's first CI workflow. Phase-0 manifest + docs-prose gates run before npm install so a broken manifest fails fast without spinning up Node; then npm ci + npm test + npm run compile.

Verification (all green locally)

• make check-manifest → dist/repo.meta.json valid; all exposes.* present ✓
• make check-docs-prose → no docs/ directory ✓
• Canonical Track-A validator (.github/profile/build/validate-repo-meta.py) → OK: dist/repo.meta.json

Design notes

• Hand-authored dist/extension-info.json (not generated). The values it mirrors (version, publisher, engine pin, settings schema, marketplace id) live in package.json. Generating would mean parsing package.json to emit a re-shaped subset — for a four-key settings block that ships once per release, a hand-authored mirror with an AGENTS.md guardrail ("update in the same commit") is cheaper than a generator. Same pattern m-test-engine uses for dist/lifecycle.json.
• m-stdlib has architectural priority over m-stdlib-vscode. Captured in AGENTS.md § Guardrails: if a needed manifest field is missing, propose adding it to m-stdlib first; do not synthesise it client-side.
• Manifest-discovery order is contract. The four-step order (setting → $M_CLI_MANIFEST → workspace walk-up → bundled assets) is referenced by m-stdlib's docs and by m-cli's --manifest flag story. Reordering = breaking change. Captured in AGENTS.md § Guardrails and pinned in dist/extension-info.json.
• Manifest gate before npm install. Cheaper to fail in 100ms on malformed JSON than to wait for npm ci + tsc. The CI ordering reflects this.

Surprises

• None on .gitignore — unlike the sibling tree-sitter-m-vscode case the planning doc warned about, this repo's .gitignore does not ignore dist/. It only ignores node_modules/, out/, *.vsix, .vscode-test/, *.tsbuildinfo, .DS_Store. So no !dist/repo.meta.json / !dist/extension-info.json exception lines were needed.
• The repo already has a remote branch origin/docs/readme-refresh carrying an unrelated v0.1 → v0.2 Status-line tweak. That change is not in this PR — restored README.md to main's HEAD before staging so the diff stays scoped to the Phase-0 contract.

Test plan

• All local gates green
• Canonical Track-A validator green
• CI green on this PR (manifest + docs-prose gates run before npm ci, then npm test + npm run compile)