A high-water mark has no memory. `swept=454/666` answers how far one pass travelled before the clock stopped it [4]. It cannot answer how far that pass travelled beyond where an earlier pass already went, because every run starts at index 1 and walks the same alphabetical order [6].
So do the subtraction the metric leaves out. The second day published 328 against the first day's 302, a gain of 26 names [1]. At that pass's own measured rate of 158.8 seconds per tool, re-covering the first 302 consumed 47,958 seconds, which is 13.3 hours of a 14.5-hour pass [2]. Roughly seventy minutes of that night went anywhere new [2]. The third day did better in absolute terms, 126 new names, and still spent 12.3 of 17.0 hours on ground already swept [3]. Published coverage rose from 45.7% to 68.2% [5] while the frontier crept forward by a couple of dozen tools a night.
The calibration is the interesting failure, and the numbers are one operator's own logs from one box, regraded by day [4]. Sizing a timeout at nearly double a measured full pass is defensible practice, and it is only as durable as the two inputs behind that measurement: per-tool cost and corpus size [14]. Both moved. The three observed days average 147.4 seconds per tool, about 2.8x the rate the bound was sized against, and the corpus now stands at 673 [15][16]. Spend the same 72,000 seconds at the new rate and it buys 488 tools, 72.5% of the corpus [6]. Both the corpus and the per-tool cost grew, and a constant has no way to notice either change [15].
The whole bound family is closed by multiplication. 673 tools at 147.4 seconds is 99,200 seconds, or 27.6 hours, against a 24-hour cadence [16]. A bound set exactly at the cadence still covers only 586 tools, 87.1% [4]. A bound above the cadence means the next fire lands while the previous run is still going [17].
The instrument needs another field. Three numbers per pass would have made this visible on the second morning: the index the walk started at, the index it reached, and the gain over the previous high-water. A gain of 26 against a 661-tool corpus reads as a stall with no knowledge of the alphabet required. The count on its own was true and inert.
The shape of the gap is the part that should worry an operator. Name ordering tracks function, because tools that do one job get named alike [12]. Past the cut point sit 13 voice tools, 12 ambient-listening tools, 11 operator-comms tools and 10 link-sense tools, plus the proxy, perception-fusion, egress-census and tunnel-health lanes [11]. A truncated alphabetical walk loses a subsystem rather than a sample [12]. `mesh-test-forgery` sorts there too, so the forgery detector had never run its own check [13], the same gap in scope that left every other tool past the cut point unswept. The writeup breaks off before naming a fix [18]. The families the arithmetic leaves standing are visible enough: shrink per-tool cost, shard the corpus across nights, or persist a cursor so tomorrow starts where tonight stopped.
