create --type addition
Mid-turn agent output emits IO frames with inline payload (operator ruling 2026-08-28)
OPERATOR RULING (2026-08-28, recorded releases#234 comment 5458741388): every IO frame whose payload fits the cap carries it inline; mid-turn agent output must not reach frame consumers as a bare seq pointer when the content would have fit — a consumer handed only the pointer pays an avoidable `spt endpoint digest` call.

Shipped v0.64.0 emits NO frame for mid-turn output (digest-resident only; AGENT_OUTPUT is end-of-turn by definition). This request adds the framed mid-turn emission under the existing inline-first contract: inline when short, 16KB-class cap + truncated flag + digest seq pointer past it (#22 ruling 7 unchanged).

Scoped OUT of #234 (doyle, 2026-08-28): W3 is a reader; this changes an emitter. Thin lanes, separate gates.

Shape points:
1. Kind vs attribute — OPERATOR CALL. Property measured by todlando: the #234 reader skips unknown kinds, so a NEW kind is version-safe (older cores skip, no handshake) but consumers miss mid-turn until opt-in; an attribute on AGENT_OUTPUT is day-one visible but touches the ratified taxonomy row's meaning.
2. Granularity: per digest entry (the durable quantum seq already points at).
3. Hybrid dedup respec + doc sweep ride this lane: frames.md mid-turn paragraph ("digest is where a consumer goes"), patterns.md dedup note, integration-checklist "ONE payload-carrying call per turn" row (published, perri blind-builds from it).
4. Rate re-derivation rides this lane: todlando's W3 1000-row retention bound was sized on turn boundaries (~130 events/hr peak, 1117-transcript corpus); per-digest-entry emission moves peak to ~540/hr (~4x, 1000 rows ≈ <2h peak). The lane that moves the rate corrects the number.