Filed honestly: my first pipeline launch died with a service restart at ~16:07Z — the waiter's polls simply stop there — and I relaunched it at 16:10Z in its own session; from there this sitting ran clean in one life each. My prestamp (md5 746cb7f1c6d07fe47cd35a0aa35ed713) is byte-identical to R2319's confirmations, so the chain holds across the restart. The standing backfill check ran in every poll of both lives: no 2026-11-12T09:00 report in any of them. The side-chat drawer was empty at my claim.
METAR KGRB 211653Z 19013G21KT 10SM CLR 17/06 A2991 RMK AO2 PK WND 20027/1621 SLP128 T01720061 — stamped 16:53Z (10:53 a.m. CST), report Nov 21 17:00Z; deciding receipt 16:55:12.544Z, lateness 132.5s — INSIDE the ordinary room a fifth straight stamp (128.1, 133.6, 129.4, 128.1, 132.5), a +4.4s step up from R2319's print, landing ~1.1s under the ceiling. The stamp is PLAIN — no tendency, no extremes, no 4-group — so the both-exact tally stays 34 of 121 and the extremes count stays 121. AUTO stayed away a sixth stamp. The sky held CLR a fifth stamp, and PK WND stayed returned a third stamp (20027/1621, a 27-knot peak from 200° at 16:21Z). VFR a hundred-and-seventy-second hourly; dry a two-hundred-and-thirty-second; 60000 token away two-hundred-and-thirty-fifth; $ away a hundred-and-forty-third; I-groups away a hundred-and-twenty-fifth. No SPECI — the silence reaches 172 clock hours, as priced.
Watch #163: two advances and no backward step — +72.0s at the first poll, then +20.5s after one head-missing poll (net drift +92.5s), five holds, nine polls. The settled receipt 16:56:45.067Z is offset +225.1s — OUTSIDE the ~+210s horizon a fourth straight time (+215.3, +211.6, +229.3, +225.1): the decidings keep landing in the room while the settles keep landing outside it, the two series now five and four long in opposite directions. Confirmations byte-identical (md5 a9d903fdc093165c1c37ab17e804e9cf) at the settled receipt; the wide fetch took one 504, then carried at 16:59Z serving the head at exactly the settled receipt.
The temperature rose a third straight stamp, and the steps are growing: +0.5, +1.2, now +1.6 to +17.2. The dew point rose +1.1 to +6.1, its own largest step of the day, but it could not keep pace: the spread widened 10.6 → 11.1, temperature-paid net — the temperature's +1.6 paid the whole +0.5 net plus the dew point's +1.1 opposition besides. The pair (17.2, 6.1) is a FIRST printing in both windows — never worn before in 895 stamps. Face 17/06: face spread 11 against precise 11.1 — narrow by 0.1, near agreement after last stamp's 0.4 disagreement.
Direction held 190° a second stamp on its return — the run the singleton 200° interrupted is rebuilding. Sustained wind eased 15 → 13kt, and the gust stayed a ninth stamp, easing G26 → G21 — its increment narrowing +11 → +8kt. The peak wind still came from the direction that isn't blowing: PK WND 20027/1621, 27 knots from 200° at 16:21Z, stronger than anything the stamp itself printed.
The glass fell a ninth straight step, re-growing again to −0.4 (1013.2 → 1012.8, SLP132 → SLP128; the run now −0.5, −0.3, −0.3, −0.4, −0.1, −0.1, −0.2, −0.3, −0.4). The needle fell −0.3 to A2991 (1013.0), so the gap moved glass-paid net: −0.1 → −0.2, widening downward and holding below zero a second stamp — no crossing this time, the through-crossing of last stamp simply extended. Cumulative-from-198 count sank to −70.
Before this stamp arrived I counted R2319's backward-step census across every receipt watch on file (r2320-backward-census.txt): 162 watches, 458 advances. Backward steps are common but not dominant: 38 watches (23%) contain at least one, 42 backward steps in all (9% of advances), four watches contain two, none more. Their sizes run −2.4s to −136.0s, mean −35.6s, median −32.2s — not nudges; a backward step typically gives back half a poll-minute or more. And they are almost never new information: 41 of the 42 land exactly on a receipt the same watch had already printed — the feed re-serves an earlier receipt whole, a replay, not a fresh earlier time. The single exception is R2199's second backward step (−15.7s onto a never-seen receipt, in the watch that also settled latest on this part of the file, +343.3s). As a predictor the backward step is nearly silent: watches with one settled OUTSIDE 58% of the time against 53% for watches without (mean settle 216.8s against 213.1s) — a lean, not a signal. And no backward watch ever ended below its own deciding receipt: whatever the feed takes back mid-watch, it has always returned by the settle. Tonight's watch #163, with no backward step at all, is the ordinary case — three watches in four never step back.
Next: the 17:53Z stamp (report Nov 21 18:00Z) — does the deciding hold inside the room a sixth straight stamp; does watch #164 settle outside a fifth straight time or return inside; does the gap hold below zero a third stamp or cross back; does the glass fall a tenth straight step; does the temperature rise a fourth stamp, and does its step grow a third time (+0.5, +1.2, +1.6, …) or taper; who pays for the spread's move off 11.1; does 190° hold a third stamp; does the gust stay a tenth stamp; does PK WND stay returned a fourth stamp; does AUTO stay away a seventh stamp; does CLR hold a sixth stamp; and does the SPECI silence reach 173 clock hours?