Add RegistrySync SSE transport and resilience tests

[bokelley]

Context

The registry is adding GET /api/registry/feed/stream as the Server-Sent Events transport for cursor-based registry feed pages. GET /api/registry/feed also now includes required freshness metadata.

Registry PR: adcontextprotocol/adcp#5733

SDK work

• Regenerate/import registry OpenAPI types so FeedResponse includes freshness.
• Add a stream client for /api/registry/feed/stream; generated OpenAPI types alone are not enough because the response is text/event-stream.
• Update RegistrySync to default to auto: try SSE first, fall back to polling /api/registry/feed on unsupported transport, 404/406, proxy/network failure, or stream parse failure.
• Persist cursor exactly as today; reconnect with last persisted cursor after disconnect.
• On cursor_expired, re-bootstrap then resume.
• Treat cursors as scoped to the same logical types subscription.
• Persist data.cursor from SSE payload; registry does not use SSE id: / Last-Event-ID in 3.x.

Tests

• Mock SSE server emits feed, heartbeat, disconnect, cursor_expired, and malformed data.
• Cursor advances only on feed, not heartbeat.
• Reconnect resumes from persisted cursor.
• Stream errors and unsupported endpoint fall back to polling.
• cursor_expired triggers re-bootstrap.
• Duplicate or partial pages do not corrupt indexes.
• Freshness metadata is exposed or observable for lag monitoring.

[bokelley]

Triage

Classification: Feature request
Bucket(s): library, schema-sync
Status: ready-for-human

What the experts said:

• ad-tech-protocol-expert: SSE-first/poll-fallback is sound, but SSE auth, backoff, and cursor↔types binding are unspecified; use a vetted SSE parser, not hand-rolled parsing.
• code-reviewer: split into two PRs (freshness typing vs. SSE transport); getFeed() never sets cursor_expired on 410 today, so this issue's re-bootstrap trigger is currently dead code.
• dx-expert: default transport to 'poll', not 'auto' — silently opening new connections for every caller is a real behavior change; add activeTransport/transportChange for observability.

My take: Direction's sound, but three judgment calls need your sign-off first. @bokelley:

1. transport default: 'auto' (per issue) or 'poll' (no behavior change, promote later)?
2. Cursor↔types binding: server-enforced, or must client persist them as a pair (CursorStore only keeps a bare cursor today)?
3. One PR gated on adcp#5733 merging, or split freshness-typing now / SSE later?

Bonus find: getFeed() never actually sets cursor_expired on 410 — worth its own fix regardless of this issue.

---

Triaged by Claude Code. Session: https://claude.ai/code/session_01ASCy7VNzCvACUrewDTRmHN

---

Generated by Claude Code