Stream carrier for attachments: blob chunks ride moss streams with relay and room-wire fallback

[ForeverInLaw]

Parent

#2

What to build

Attachments ride moss streams: the attachment blob channel keeps its chunk protocol (requests, retry, dedup — mid-transfer wobble still recovers) and changes only its carrier. The moss stream API gives a direct-session fast path (no discovery, no overlay stall on a hot loop) with a wrapped-relay fallback inside the library for peers without a direct leg, plus a fallback to the room wire for counterpart libraries that predate streams. User-visible effect: large files arrive noticeably sooner on direct sessions; relayed and mixed-version paths behave exactly as today.

The carrier swap must not touch the chunk protocol or its retry bookkeeping — it is a wire change behind an existing mini-protocol, and the measure-first note from the spec applies: this ticket is where the stream technique gets proven before any decision about directed sends for message data.

Acceptance criteria

• Chunk traffic rides the stream path when a direct session exists; wrapped-relay fallback carries it otherwise
• The chunk protocol is unchanged: request/retry/dedup behavior identical behind the new carrier
• Mixed version: a counterpart library without streams falls back to the room wire and transfers complete
• No pubsub room carries chunk traffic that the stream path has taken over (the room wire is fallback, not duplicate)
• Transfer tests run the chunk protocol over the new carrier against the real dlopened library
• A stand probe-e2e attachment round demonstrates transfer on both a direct and a relayed path
• Existing suites stay green

Blocked by

None (can start immediately).

[ForeverInLaw]

Shipped in 0.9.0 (2026-09-21): attachment chunks ride moss streams (Moss_OpenStream/Moss_SendStream, stream id 2) on direct DM sessions, with a byte-identical room-wire fallback for older counterparts — requests, retry and dedup unchanged. See CHANGELOG.md [0.9.0] Added, 'spec #8'.