I calibrated the final Everett-named optimized merge on an Apple M2 Max. The frozen rule selects GPU when the two inputs contain more than 51,200 records in total, provided every header feature is inside the measured envelope and the device, backend and execution fingerprint match. Otherwise it selects CPU.

The threshold is a fitted split between sampled workloads, not a measured sharp crossover. Although this training set chose record count for its single decision, compressed payload bytes, file bytes, byte sizes per record, size imbalance and value-width metadata still constrain the admitted envelope. No record walk, key reconstruction or EF selection is needed to choose a backend.

The rule selected five of sixteen held-out cases. All five passed the predeclared margin: even the slowest GPU repetition was more than 10% faster than the fastest CPU repetition. Their median speedups range from 2.03× to 5.79×. One additional robust GPU win was conservatively left on CPU. This is evidence for this workload sample, not a guarantee for every input inside the numeric envelope.

Frozen method and identity

Source revision: 6c1bbd2e1e1bb364389dd9f2b6977caf02a92527; the final renamed optional build passed all 13 CTests before collection.
GPU/backend: Apple M2 Max / Metal; CPU Apple M2 Max; macOS 26.6.2. The observed GPU registry token is retained in the calibration identity and is not assumed stable across boots.
Host SHA-256: 8206e2fd4a49e91438a752267177ddb8c26b599a52dc4d62eb8f25a5c4e827b2.
Shader-library SHA-256: 1af9ecde98985c57d861be3a6b261fdcf34269e2093131f6986a758ae6574763.
Fifty checked processes in fixed seeded shuffled order: 34 training and 16 held-out cases, one verified warm-up plus three alternating CPU/GPU measured pairs per case. All 150 measured pairs and 50 warm-up pairs passed complete canonical file equality, semantic scanning and CRC checks. All processes exited successfully; artifact hashes were unchanged after collection.
The split, maximum tree depth, minimum GPU-leaf population and 10% margin were fixed before measurement. The training-only rule was written and hashed before held-out scoring. Held-out results did not change a threshold or prune a branch.
Both complete totals start with existing warm compressed mapped inputs and finish with checksummed mapped output. They include resource allocation, GPU parse/tree/cache/order/count/scan/emit/output-EF work, host waits, framing and CRC. They exclude fixture generation, pipeline setup, correctness verification, final local-object destruction, fsync and catalog publication. CPU uses the existing serial compressed merger and canonical encoder. These are construction measurements, not durable-write throughput.

The generated header's decisions match the Python model on all 50 cases in a strict C++20 compile/run; all 50 wrong-identity checks select CPU. An unmatched binary, driver/OS fingerprint, GPU or backend, failed calibration validation, or features outside the training envelope selects CPU. The caller must first establish the supported sort/profile and immutable input ownership; this header-only selector does not validate a schema. The execution backend retains its independent grammar and resource checks.

Measured envelope

All sums below combine the two inputs. Imbalance is larger divided by smaller. Fixed-value bytes describe the encoded wire width, including framing; zero means unavailable for that feature. Terminal key bytes describe only the last key, not a maximum over all keys. Exact bounds are retained in the JSON model.

Header feature	Minimum	Maximum
records	512	524288
payload_bytes	8031	126940002
file_bytes	8760	127016152
payload_bytes_per_record	10.8804702759	516.437887753
record_imbalance	1	16
byte_imbalance	1	15.9983369639
terminal_key_bytes	8	264
fixed_value_bytes	0	514.5
known_value_widths	0	2

All 50 case summaries

Times are milliseconds, median [minimum–maximum] over three measured pairs. N and ratio describe the larger input and larger/smaller record counts. The decision uses combined header counts and the full envelope, not the case labels. “Robust” is the predeclared slowest-GPU versus fastest-CPU margin.

Case	Split	N	Ratio	Prefix bytes	Ties %	Value bytes / fixed	CPU ms	GPU ms	CPU/GPU	Choose GPU	Robust
0	train	256	1	32	50	8 / no	0.307 [0.282–0.392]	5.041 [2.648–5.206]	0.061×	False	False
1	train	1024	1	32	50	8 / no	0.643 [0.574–0.663]	4.341 [2.819–5.367]	0.148×	False	False
2	train	4096	1	32	50	8 / no	1.882 [1.843–1.891]	2.515 [2.470–2.937]	0.748×	False	False
3	train	16384	1	32	50	8 / no	6.941 [6.852–7.050]	6.384 [4.133–6.402]	1.087×	False	False
4	train	65536	1	32	50	8 / no	27.819 [27.392–28.460]	8.036 [7.931–8.620]	3.462×	True	True
5	train	262144	1	32	50	8 / no	109.157 [109.020–110.682]	10.697 [8.592–12.346]	10.204×	True	True
6	train	4096	4	32	50	8 / no	1.257 [1.254–1.345]	3.198 [3.033–3.444]	0.393×	False	False
7	train	4096	16	32	50	8 / no	1.168 [1.105–1.197]	3.554 [3.261–3.875]	0.329×	False	False
8	train	4096	1	0	50	8 / no	1.757 [1.744–1.806]	4.181 [2.747–4.330]	0.420×	False	False
9	train	4096	1	256	50	8 / no	1.814 [1.790–1.882]	4.370 [4.078–6.758]	0.415×	False	False
10	train	4096	1	32	0	8 / no	2.151 [2.118–2.276]	4.940 [4.262–4.954]	0.435×	False	False
11	train	4096	1	32	100	8 / no	1.477 [1.453–1.531]	6.225 [5.845–6.284]	0.237×	False	False
12	train	4096	1	32	50	64 / no	2.082 [1.961–2.244]	5.033 [4.249–5.909]	0.414×	False	False
13	train	4096	1	32	50	512 / no	3.051 [3.021–3.176]	4.603 [3.978–4.756]	0.663×	False	False
14	train	4096	1	32	50	8 / yes	1.667 [1.623–1.786]	4.690 [3.990–5.818]	0.356×	False	False
15	train	65536	4	32	50	8 / no	18.284 [17.839–18.813]	5.407 [2.863–7.814]	3.382×	True	True
16	train	65536	16	32	50	8 / no	15.943 [15.639–15.966]	6.336 [4.008–7.195]	2.516×	True	True
17	train	65536	1	0	50	8 / no	27.097 [26.799–27.815]	6.955 [4.949–7.548]	3.896×	True	True
18	train	65536	1	256	50	8 / no	27.832 [27.308–27.940]	8.959 [6.876–9.683]	3.107×	True	True
19	train	65536	1	32	0	8 / no	32.398 [31.573–32.548]	7.206 [4.395–8.051]	4.496×	True	True
20	train	65536	1	32	100	8 / no	20.173 [20.037–20.275]	5.603 [5.322–6.137]	3.600×	True	True
21	train	65536	1	32	50	64 / no	28.946 [28.682–32.891]	6.290 [5.301–8.877]	4.602×	True	True
22	train	65536	1	32	50	512 / no	45.242 [44.914–46.821]	12.407 [12.352–12.761]	3.646×	True	True
23	train	65536	1	32	50	8 / yes	24.345 [24.213–24.498]	7.814 [5.060–8.486]	3.116×	True	True
24	train	262144	4	32	50	8 / no	73.592 [73.295–73.678]	6.913 [5.640–10.522]	10.645×	True	True
25	train	262144	16	32	50	8 / no	64.261 [63.748–64.287]	9.821 [7.559–9.967]	6.543×	True	True
26	train	262144	1	0	50	8 / no	110.462 [109.722–111.339]	10.723 [8.438–11.027]	10.301×	True	True
27	train	262144	1	256	50	8 / no	110.437 [109.896–112.393]	12.968 [11.494–16.395]	8.516×	True	True
28	train	262144	1	32	0	8 / no	131.703 [131.034–132.256]	8.803 [8.775–12.564]	14.961×	True	True
29	train	262144	1	32	100	8 / no	82.313 [81.162–84.032]	7.257 [6.473–7.911]	11.343×	True	True
30	train	262144	1	32	50	64 / no	118.719 [116.965–119.226]	12.105 [10.581–19.128]	9.807×	True	True
31	train	262144	1	32	50	256 / no	145.438 [144.048–145.781]	21.340 [20.415–30.980]	6.815×	True	True
32	train	262144	1	32	50	8 / yes	98.151 [97.452–99.088]	8.370 [7.597–11.426]	11.726×	True	True
33	train	65536	16	256	0	512 / yes	28.275 [27.462–29.576]	19.946 [16.502–20.227]	1.418×	True	True
34	heldout	2048	1	64	25	32 / no	1.173 [1.079–1.216]	6.052 [5.458–6.637]	0.194×	False	False
35	heldout	2048	3	128	75	128 / yes	0.708 [0.686–0.768]	4.376 [3.473–5.666]	0.162×	False	False
36	heldout	2048	8	0	0	512 / no	1.158 [1.125–1.426]	2.684 [2.541–3.267]	0.431×	False	False
37	heldout	2048	16	256	100	8 / yes	0.588 [0.581–0.641]	5.643 [4.977–5.872]	0.104×	False	False
38	heldout	8192	1	64	25	32 / no	3.989 [3.792–4.054]	6.288 [4.762–6.327]	0.634×	False	False
39	heldout	8192	3	128	75	128 / yes	2.700 [2.495–2.832]	5.355 [3.842–5.946]	0.504×	False	False
40	heldout	8192	8	0	0	512 / no	4.328 [4.253–4.361]	4.605 [4.301–4.847]	0.940×	False	False
41	heldout	8192	16	256	100	8 / yes	1.911 [1.871–2.151]	5.772 [5.442–6.379]	0.331×	False	False
42	heldout	32768	1	64	25	32 / no	15.653 [15.646–15.728]	7.692 [4.124–8.381]	2.035×	True	True
43	heldout	32768	3	128	75	128 / yes	10.242 [9.830–10.484]	8.945 [5.811–11.288]	1.145×	False	False
44	heldout	32768	8	0	0	512 / no	16.191 [15.597–17.908]	8.999 [6.509–9.051]	1.799×	False	True
45	heldout	32768	16	256	100	8 / yes	6.769 [6.499–6.800]	7.188 [6.497–7.788]	0.942×	False	False
46	heldout	131072	1	64	25	32 / no	63.666 [62.399–64.479]	11.001 [8.795–12.840]	5.787×	True	True
47	heldout	131072	3	128	75	128 / yes	39.707 [37.944–40.386]	11.770 [10.123–12.623]	3.374×	True	True
48	heldout	131072	8	0	0	512 / no	61.858 [61.639–66.110]	17.417 [15.845–19.251]	3.552×	True	True
49	heldout	131072	16	256	100	8 / yes	26.997 [26.574–27.713]	7.214 [5.907–13.484]	3.742×	True	True

Artifacts and use

The optional selector is header-only. The generated calibration header records this exact identity, envelope and tree. It is a conservative construction-backend hint; it does not install a GPU backend into the durable runtime. Include the optional directory in the compiler include path when using the generated header.

All raw trial rows, all case summaries, the frozen training rule, the exact calibration, and the per-case raw CSVs remain available. Qualification records the new named build separately from original public normalization. Provenance preserves original and public metadata hashes; public metadata omits local paths and hostname, while timing rows, artifact fingerprints and model values are unchanged.