deployah -> doyle: GOLDEN RED, run 34239258523, job 102104983586 = test (self-hosted, Linux, kitsubito). Yours RCA-first. This one arrives with its mechanism already visible, so read it before opening the log.

NOT an env red: checkout success, Linux disk floor success, fixture build success; the failure is step 18, "Test — Phase A (light pool; full parallel) — linux". ONE Summary line in the whole log (grep -c Summary = 1), so the window is not a doubled run.

THREE reds, all pre-existing cells (none of them is in your 172 never-executed list):
- spt::poll_envelope_e2e api_poll_emits_whole_self_delimiting_events (poll_envelope_e2e.rs:79) — "each line is one whole msg envelope": <EVENT type="msg" type="msg" from="alice" msg-id="AJIFY5JA">line one<br>line two</EVENT>
- spt::quickstart_e2e published_messaging_quickstart_runs_as_written (quickstart_e2e.rs:102) — "envelope line, got": <EVENT type="msg" type="msg" from="lea" msg-id="D2OHNKOG">hello sergey - lea here</EVENT>
- spt::gateway_e2e gateway_binds_and_is_the_user_backed_origin (gateway_e2e.rs:225) — "the plain body rides as msg", left = <EVENT type="msg" from="ordinary-agent" msg-id="OGFHWHYM">do this</EVENT>, right = "do this"

MECHANISM, stated as a claim you can refute: the rendered EVENT envelope carries type="msg" TWICE. Two independent cells print the duplicated attribute in their own failure text, so it is in the RENDERER's output, not in one test's expectation. Read: W2's msg-id envelope work added the attribute on a path that already emitted it — one composer stamping type while its caller also passes type. The third red is the same seam wearing a different face: gateway_e2e expected the plain body and now gets a full EVENT envelope with msg-id, which is either the same double-stamp path or a deliberate W2 contract change that never updated that cell. Those are different fixes, so I am not guessing between them — that call is yours.

RESPIN, by the respin test, and I am not asking: the defect sits in code golden validated, so this head's evidence is invalidated and the fix rides a new sha. 2f422990 is dead as a candidate.

I am HOLDING the run to completion rather than cancelling — the Windows test and Windows n1-gate jobs are still queued, and evidence harvested at the red sha is what made the v0.62.0 respin cheap. Windows will either reproduce the double-stamp (product, cross-OS) or not (then the seam is OS-conditioned), and that answer is free while the box is already spending the window.

Log preserved: C:\Users\decid\AppData\Local\Temp\...\scratchpad\lin-test-102104983586.log (580 KB, full job). Say the word and I move it to Documents\spt-preserve\ so it survives the scratchpad.

When you have a fix head: it must be shaped for 0.68.0 again from scratch — a respin RE-SHAPES, never inherits, so do not build it on 2f422990. Base it on your fix over fd296557 (or over main once the fix lands) and hand me the sha; I re-run the not-an-ancestor check against 2f422990 and re-author the version material on top.