Add Go SDK (P1.1) — v0.2.0 surface parity

[jaschadub]

Summary

Brings AgentPin to four-language parity by adding a Go SDK at
github.com/ThirdKeyAi/agentpin/go, mirroring the v0.2.0 stable surface
of the Rust crate. This addresses the P1.1 priority from the cross-stack
gap analysis: a Go SDK is the missing fourth language alongside Rust,
JavaScript, and Python.

Scope

v0.2.0 surface only. The two in-flight v0.3.0-alpha.1 features
(A2A AgentCard types and DNS TXT cross-verification, currently in PRs
#2 and #3) are intentionally NOT included in this PR. They will
follow in a separate Go alpha PR after the corresponding Rust PRs
merge — keeping this change focused on bringing Go to v0.2.0 parity.

Package layout

Mirrors SchemaPin's Go SDK conventions:

go/
├── go.mod module github.com/ThirdKeyAi/agentpin/go
├── README.md install, quickstart, API map
├── cmd/agentpin/ CLI: keygen | issue | verify | bundle
├── internal/version/version.go declares 0.2.0 (matches Rust/JS/Python)
└── pkg/
├── crypto/ ECDSA P-256 keypair + DER signature helpers
├── jwk/ PEM ↔ JWK + RFC 7638 thumbprint
├── jwt/ ES256-only JWT (rejects everything else)
├── types/ Shared types (capability, constraint, discovery, etc.)
├── discovery/ Build / validate / fetch discovery documents
├── credential/ IssueCredential + capability subset validation
├── verification/ 12-step VerifyCredentialOffline + resolver variant
├── revocation/ Build / check / fetch revocation documents
├── pinning/ TOFU KeyPinStore with JSON persistence
├── delegation/ Attestation create / verify + chain depth check
├── mutual/ 128-bit nonce challenge / response
├── nonce/ Replay-protection store interface + in-memory impl
├── bundle/ TrustBundle helpers
└── resolver/ WellKnown / LocalFile / TrustBundle / Chain resolvers

Critical security requirements (all met)

• ES256-only enforcement.jwt.DecodeJWTUnverified rejects every
  algorithm except ES256 and every typ except agentpin-credential+jwt
  before any signature work. We use crypto/ecdsa directly — no
  third-party JWT dependency with permissive alg defaults. Tests
  cover none, HS256, RS256, ES384, and empty-alg cases.
• Wire format compatibility. RFC 7638 JWK thumbprint, discovery
  documents, credentials, revocation lists, and trust bundles all
  round-trip byte-identically with the Rust SDK. Asserted by
  cross-language interop tests (see below).
• 12-step verification flow preserved verbatim from the Rust crate
  (documented in the package doc comment).
• No eval/dynamic code. Rate-limit and domain-pattern matchers
  use closed-vocabulary parsers.

Cross-language interop test results

go/pkg/verification/cross_language_test.go exercises every wire
format boundary using fixtures generated by the Rust CLI:

Test	Result
TestCrossLang_PEMtoJWK_MatchesRust	PASS
TestCrossLang_JWKThumbprint_MatchesRustPort	PASS
TestCrossLang_DiscoveryDocument_RustSerialization_RoundTrip	PASS
TestCrossLang_RustKey_GoSignVerify	PASS
TestCrossLang_JWT_RustGenerated_GoDecodeAndSignatureVerify	PASS

End-to-end manual interop test also confirmed that the Rust CLI
verifies a Go-issued credential (and vice versa), proving the JWT
DER-signature wire format is byte-compatible across SDKs.

QA matrix

Check	Status
go test ./...	PASS — 106 tests across 13 packages, 0 failures
go vet ./...	clean
gofmt -l .	empty (every file gofmt'd)
cargo test --workspace	PASS — 135 Rust tests still pass
cargo clippy --workspace	clean (no warnings)
Bidirectional CLI interop	Rust verifies Go-issued cred and vice versa

Test count by package

Package	Tests
pkg/types	23
pkg/crypto	5
pkg/jwk	5
pkg/jwt	11
pkg/discovery	7
pkg/credential	2
pkg/verification	18 (incl. 5 cross-language)
pkg/revocation	5
pkg/pinning	6
pkg/delegation	4
pkg/mutual	6
pkg/nonce	3
pkg/bundle	3
pkg/resolver	8
Total	106

CI

• New .github/workflows/go.yml runs gofmt -l, go vet, and
  go test across Go 1.21 and 1.22 on every PR touching go/**.
• The version-consistency job in .github/workflows/release.yml is
  extended to also validate go/internal/version/version.go so the
  four manifests stay in lockstep.

Documentation

• go/README.md — install (go install …@latest), quickstart, API
  reference table mapping Go symbols to the Rust equivalents, security
  guarantees, and a manual fixture-regeneration recipe.
• Top-level README.md — Go added to the SDK list and project structure.
• SKILL.md — Go quickstart section in the same shape as the existing
  JS/Python sections; language API reference table extended.
• context7.json — description and folders updated; *.go added to
  excludeFiles.
• CHANGELOG.md — Unreleased section documenting the new SDK above
  the existing 0.2.0 entry. No existing entries were changed.

Deviations from the Rust API

Minor, idiomatic-Go surface-only changes:

• Verification result fields use Go-style PascalCase (AgentID vs
  Rust agent_id) but JSON tags preserve the snake_case wire format.
• KeyPinStore is goroutine-safe via an internal sync.Mutex
  (Rust's struct is !Sync; callers wrap with Mutex themselves).
• The WellKnownResolver exposes an injectable *http.Client field
  so callers can plug in their own transport; Rust uses a private
  reqwest::Client built per-call.
• All other functions, types, JSON field names, and error codes match
  the Rust API symbol-for-symbol.

Follow-ups (not in this PR)

• v0.3.0-alpha A2A + DNS TXT for Go. Will land in a follow-up PR
  once the Rust PRs v0.3.0-alpha.1 (P0): A2A AgentCard types + AllowedDomains (Rust) #2 (A2A AgentCard) and v0.3.0-alpha.1 (P1.2): DNS TXT cross-verification at _agentpin.{domain} (Rust) #3 (DNS TXT
  cross-verification) merge. The Go SDK directory layout already
  leaves room for a pkg/dns/ package.

Pre-existing CI infrastructure issues

These are not introduced by this PR but may show up red on the CI
run; flagging for transparency:

• The Rust CI matrix pins MSRV to 1.70, which currently fails with a
  clap_builder edition-2024 transitive dep error on dependency
  resolution. This is independent of the Go SDK and would be in scope
  for a separate dependency-pin PR.
• The JavaScript CI workflow runs npm ci, but package-lock.json
  is gitignored at the repo root, so npm ci cannot succeed without
  generating a lockfile. Also out of scope for this PR.

Test plan

• cd go && go test ./... — 106 tests pass
• cd go && go vet ./... — clean
• cd go && gofmt -l . | grep -v '^$' — empty
• Cross-language interop tests pass (5 tests)
• cargo test --workspace -j2 — Rust 135 tests still pass
• cargo clippy --workspace -j2 -- -D warnings — clean
• Bidirectional CLI interop verified manually (Rust verifies Go
  credential; Go verifies Rust credential)