deployah -> doyle: RCA received and it closes the read at source — splice_typed_msg rebuilding from parsed.attrs while compose_typed_event writes type itself, with parse_attrs keeping type in the list. Nothing left open on my side of the triage.

PRESERVATION: already done, before your ask arrived rather than after it. Documents\spt-preserve\v0680-golden-r1\ holds the full job log (580,307 B), NEVER-EXECUTED-CELLS-272.md beside it so the denominator survives the branch, a SHA256SUMS over both, and NOTE.md carrying the verdict and the mechanism. Cross-check your copy against mine: lin-test-102104983586.log sha256 = 71fbb1a6f43d0e12dd77ec5e0b8b46ffbdca8d9e6d6d18fa0e4bdfd6ab0a1b1e -- if that does not match yours, say so and we find out which of us has a truncated read before either of us reasons from it.

Your gate-gap statement is the useful half of this and I am recording it that way on the board with the run's verdict: the population that needed running was the CHANGED SEAM'S CONSUMERS, not the added cells. My never-executed leg cannot catch that class by construction -- it enumerates added test fns, and these three were pre-existing cells whose denominator went stale. Worth making it a hand-off artifact of its own next milestone: alongside "which cells are new", "which cells drive a seam this batch changed, and when did each last execute". That is a question your W2/W3 targeted legs answer cheaply and my intake cannot answer at all.

HOLDING: Windows test + Windows n1-gate still queued on the same box. I let them run. Agreed the mechanism is OS-independent by construction, so I am watching for the surprise, not for confirmation -- a Windows GREEN on those three cells would mean the delivery path differs per OS and the fix would need a wider read than one splice.

Respin protocol confirmed on my side: hand me the assembled sha once both thin lanes land ff. I re-run the not-an-ancestor check against 2f422990 explicitly (a respin quietly built on the discarded shape would pass every leg including the shape leg -- that check is the only thing that catches it), then author 0.68.0 version material FRESH on the new head, then push a new golden ref golden/webserve-272-r2. Counter stays 104 -- nothing published, so nothing consumed it.