docs: name the endpoint and streaming behavior agent.md left out

[miharp]

Why

Came up answering "what's the transport for getting the sealed artifact onto the agent, and is that well documented?" The design rationale (pull not push, resealing not checksum, mutual TLS as the model) is well covered — but the actual mechanics of the fetch step were thin, and split across files in a way that made them hard to find.

agent.md's reconciliation list — the doc specifically about the compiler side — just said "Fetch the artifact, extract it to a temporary directory." No endpoint, no mention it rides the same mutual TLS as the poll one line above it. The endpoint (GET /v1/artifact/{environment}/{code_id}) is documented, but only in publishing.md, written from the publisher's side. And two mechanical details weren't written down anywhere at all:

• The response body is piped directly into the extractor — no intermediate file on disk.
• Content-Type: application/gzip is the archive format, not an HTTP Content-Encoding the transport adds.

What changed

• Step 3 now names the endpoint and the shared mutual TLS.
• New section, "The fetch is streamed, not staged", placed right before "Verification is by resealing, not by checksum" (continuing the same fetch → verify → swap flow the rest of the doc already follows), covering both details above.

Verified against source before writing it

ExtractArchive (internal/seal/archive.go): gzip.NewReader(r) wraps the io.Reader directly, tar.NewReader reads it header-by-header via tr.Next(). No buffering step anywhere between the socket and the extracted files — confirmed this is genuinely a streaming pipeline, not just calling it one.

Checks

markdownlint-cli2 and cspell clean.

🤖 Generated with Claude Code