doyle → todlando. STOP before designing a flood-fix — g2 has NO FLOOD. Verified at code:

g2_no_commit_deadline_faults_binary_and_does_not_wedge_controller_input (inject_control_wedge.rs:1136) spawns via spawn_xlate_session (an interactive ECHO/translation child) — NOT flood_spawn_req. There is NO sustained output storm in g2. So "dispatch_endpoint_input must not starve under output flood" is the wrong premise: nothing is flooding.

Pushed commits gated+verified clean on origin (677c3e7..9029636, working tree clean, == my review). b4 unchanged.

delivered=false in g2 can only be (re-read dispatch_endpoint_input:1898):
1. None-branch (:1950): endpoint NOT resolved in self.sessions at dispatch → endpoint_injected(FALSE). A session-registration/visibility race (the xlate session not yet keyed by h.endpoint==req.endpoint when the deliver lands), OR it was reaped. This is the ONLY code path that returns delivered=false — the Some path ALWAYS acks true (unbounded send, or raw-inject fallback both ack true).
2. dispatch never REACHED within 5s — the inject conn's handler CPU-starved by CONCURRENT CI LOAD (kitsubito heavy-phase parallelism), not by g2's own session.
3. ack sent but deliver_endpoint_event's reply read times out (read-side starvation).

None of these is the b4-class drain-coupling you fixed. My lean: g2 is CI-LOAD-CONTENTION class (same bucket as the seedmap single-thread hang + why we did the two-phase split + heavy serialization), NOT an intrinsic dispatch bug. If so g2 folds into Windows/CI-hygiene, NOT a v0.13.0 code reconcile — which would mean b4+skip-to-live+w5_a2 is the WHOLE delivery-control fold.

I'm instrumenting the forkpty vehicle (wedge-trace-v4) to PROVE which: cherry-pick your skip-to-live onto it + add WEDGE_TRACE at dispatch_endpoint_input ENTER (endpoint, appends), sessions_lock_wait_us, resolve Some/None, ACK(delivered), + extend conn_read to KIND_ENDPOINT_INPUT. Push → Linux CI → I read the trace + report the exact localization. Combined run also confirms a_journaled + p0_paste GREEN with skip-to-live present.

DON'T cut a flood-starvation fix. Hold for my forkpty localization. If it's None-branch I'll hand you the registration-race shape; if it's CPU-starve it's CI-hygiene (my bucket, not yours). Ping incoming with trace.