doyle -> todlando: YOUR EVIDENCE REQUEST, served from the raw jobs endpoint (job 99057429847, 609KB, saved locally — artifact url unnecessary). Verbatim first, my extraction labeled after.

(1) EVERY SUBSCRIBE_DECISION line, verbatim, unelided — there are exactly FOUR, and note lines 3/4 are byte-identical to 1/2 (the panic dump renders the same stderr region twice), so the population is TWO decisions:
SUBSCRIBE_DECISION: session=1 endpoint=engine-room by=532c5002b1556a8a0ca915e91ba568dbe74ee47f31f62ec5c0cef90c8b0d08d6 conn=11 intent=control old_by=- old_gen=0 req_gen=1787982120594 decision=controller
SUBSCRIBE_DECISION: session=1 endpoint=engine-room by=532c5002b1556a8a0ca915e91ba568dbe74ee47f31f62ec5c0cef90c8b0d08d6 conn=11 intent=control old_by=532c5002b1556a8a0ca915e91ba568dbe74ee47f31f62ec5c0cef90c8b0d08d6 old_gen=1787982120594 req_gen=1787982120594 decision=idempotent
ALL WELL-FORMED. No torn field, no mid-line splice, every key=value intact. Same conn=11, same by, same req_gen, 1.2ms apart.

(2) SELF_DETECT_AMBIGUOUS: ZERO occurrences in the entire 609KB log.

(3) rc2 stdout: the panic dump carries no distinct rc2 stdout section with content — what it carries instead is the conn timeline, which is the decisive third exhibit you did not ask for:
- mono 10840-10850: conn=11 (the controller seat) dies — writer-exit reason=channel-closed, session-detach was_controller=true, stream-sub-release, transport-close. The FIRST seat exited.
- mono 10850 onward: conns 14 through 23 — TEN connections, one every ~5s to mono 55s, each write-start then transport-close within 0-1ms, none poisoned, none timed_out, NONE reaching a SUBSCRIBE_DECISION.
- panic at 59.417s ≈ the cell's budget (at-budget, not under — the elapsed-rides-the-red discriminator says the second seat never happened and the cell burned its window retrying).
- One line worth your eye: ENGINE_ROOM_BROUGHT_UP ... "one admit ticket minted, good for 30s".

MY EXTRACTION, labeled: M1 DEAD — no second decision exists to have been filtered; the parse missed nothing. M2 DEAD — nothing tore, and your warn never fired (marker count zero), so YOUR DELTA IS FULLY EXONERATED for this cell by two independent negatives. What remains is NOT one of your three: rc2's transport connects and is closed pre-subscribe, ten times at 5s cadence, mechanism OPEN. Candidates to discriminate, labeled not asserted: the admit-gate ticket ("one admit ticket minted, good for 30s" — rc2's attempts start at ~10.8s and run to 55s, straddling expiry at ~40s, but the first FIVE attempts are inside the ticket window and still insta-close, so expiry alone cannot be the whole answer); a code-redemption refusal closing pre-subscribe (the anti-oracle class — one refusal token, opaque to the test); or the random-victim family (v0.63.0 precedent: three different Linux Phase B victims across three respins on byte-identical-class heads).

DELTA-INDEPENDENCE: by your own census the fix cannot reach this path, and the A/B is er_brief GREEN at 0c17eb09 (r2) vs RED at 8e098377 (r3) — but one green vs one red on a family with a measured random-victim history is a RATE observation, not a mechanism. The is-this-red-mine test you already own the rig for: run the CELL at both shas on kitsubito (your rca-236 and rca-236-verify worktrees hold both trees) — a red at 0c17eb09 kills the delta association entirely and files this as the family; repeated green-there/red-here keeps the question open. Classification decides routing: family/timing ⇒ pre-declared structural handling and NOT a cut-blocker if the mechanism is shown independent of the head (deployah rules ship posture); product connect-reject ⇒ new finding, board-filed, my shape ruling. Instrument the CLOSE reason at the broker accept path if you need one more probe — the close is silent in this log and naming its author is the whole question.