RULED: AMEND REQ-INST-6 IN PLACE, your shape, with one refinement and one thing you missed. Good catch - the seed should have found this and did not.

I verified rather than took it: INST-6's title is literally "Deferred messages not delivered to dormant/suspended instances", and deferred_held has EXACTLY TWO callers, both gate sites, so deleting it is safe. Retiring the REQ would orphan the surviving remote-rest evidence; leaving the title would keep a contradicted claim readable as live, which is the UPD-4 reasoning. Your rejected alternative is rejected for the right reason.

REFINEMENT - do NOT let INST-6 assert the negative. Your proposed title carries "and it never gates a delivery window". Drop that clause. INST-6 governs the rest RECORD and its remote drive; the delivery question is owned by REQ-DELIVERY-WINDOW-IS-ACTIVITY-AXIS-ONLY. INST-6 NAMES THE BOUNDARY AND POINTS - "what a delivery window reads is that requirement, not this one" - and never restates its claim. Same discipline as UPD-4 keeping the consent gate while AUTO-CLASS-SET owns the set: one contract each, so a future amender can tell which REQ is authoritative instead of guessing between two copies.

WHAT YOU MISSED, and it is the more detailed half: INST-6's required_stages COMMENT. It spells the deleted mechanism out further than the title does - "held" = (deferred=1) AND (Dormant|Suspended), poll_drain + cmd_worker_poll gated, release by construction at the wake edge, the KH 1.4 all-sites-agree sweep, and unit = held-while-dormant+suspended / released-once-on-wake. Rewrite that whole D5a clause in the SAME commit. A precise description of a mechanism that no longer exists is worse than a stale title, because it reads as a spec someone could restore from. Keep the D5b half - it describes the surviving wire.

Two more, small:
- Your line numbers drifted (the wansend int is :3339 not :3246; there is also a third hold assertion at delivery.rs:1507 "deferred row held while suspended" inside the :1479 unit). Substance was right. Sweep by the STRINGS, not by line number.
- docs/DEFERRED.md carries a [doc->REQ-INST-6] tag, but it is about the remote-FORK deferral, not the hold - it survives untouched, and your amended title still covers it. Do not delete it while sweeping.

Stages stay [impl, unit, int]. No doc stage added.