Composition, not coherence, was the confound worth testing, and length survives it¶

Two explanations for Qt's permutation moving from p = 0.238 at 64 examples a client to the 1/84 floor at 128 were tested and failed: pairwise coherence (Qt stays at -0.0005 at 128) and, post hoc, centred centroid coherence (Qt is positive at both lengths and lower at 128, +0.109 to +0.047). Rather than generate a third, this tests the confound the training-length entry already named: the two sets differ in composition as well as length, because packing at 128 leaves fewer clients per project (AOSP's system-core falls from three clients to one, frameworks-av from four to two).

Five random subsamples of the 64-example clients, each drawn to the 128-example set's exact per-project counts (34 C++ clients), then the same project-level permutation:

Composition accounts for part of the movement -- matching it takes Qt from 0.238 to a median of 0.131 -- but with 64-example updates Qt reaches the floor in none of five draws, and with 128-example updates it does. The longer training, not the repacking, is what makes Qt's grouping the most extreme of all 84. AOSP is at the floor in every draw at both lengths.

The limit on this, stated: the 128-example side is one packing, since a second needs another training run; the 64-example side is five. The comparison is five draws that never reach the floor against one that does, which is strong but not symmetric. Job 149598, at 256 examples a client on the same committed source list, is the next point, and a second 128 packing would make it symmetric.

That leaves the training-length claim standing for both organizations, and leaves the mechanism for Qt open: whatever longer training adds, it is not pairwise or centroid coherence. Tamura and Tsugawa (arXiv:2609.19864, submitted 17 September 2026) offer a candidate worth measuring rather than asserting: across GitHub, identifier-naming diversity fell around 2023-2024 as LLM tools spread, yet "repositories whose owners are closer in the collaboration network remain more similar in naming style even among recent, more homogeneous cohorts", in five of six languages. If organizational coherence in update space tracks collaboration density, AOSP's platform team and Qt's looser federation of framework, IDE and runtime would differ exactly as they do. The built examples do not carry reviewer identities, but the raw snapshots carry stable salted pseudonyms, so reviewer overlap across an organization's projects is measurable without identifying anyone.

That also answers a threat to the code-convention finding. If naming styles are converging under LLM tools, an organization's convention fingerprint might be eroding. Tamura and Tsugawa find that aggregate convergence and network-local variation coexist, and an organization is a network-local cluster by construction. (A second hit, Kupari, Giacaman and Terragni, ICSME 2025, arXiv:2601.09832, was summarized elsewhere as finding code-style adherence stable over twelve months; its abstract page does not state that, so it is not cited for it.)