Source: 20efcf2af33b38c5be7596e2db7b243a3be53c47. Nearest same-input baseline: e3ec6d231cc7e5c769685761a4739c8fcffc30ff. All 51 timed observations are retained in the checkpoint data, including every full-output fingerprint. The cross-revision full-output equality check failed. None of these rows was appended to the existing benchmark history, and its equality guard remains unchanged. Rendering this diagnostic report is not a passing benchmark-promotion verdict.
Timing and output identity
Points show three-pass medians; ranges show the observed minima and maxima. Source labels distinguish the same-input baseline and diagnostic checkpoint. These are timing records, not paired speedup claims: the singleton SQL modes have unequal full-output fingerprints across revisions.
The equality grid compares recorded fingerprints within matching workloads. A matching literal-replay projection does not imply equality of the complete output contract.
Workload and execution
The benchmark driver uses the same registered 180-base canonical-DNA CDS, four original ALT events and diploid sample templates as the baseline report. It checks complete carrier multisets, sequences, edit provenance, corrupted inputs and full output fingerprints within each mode. The singleton HGVS check compares independent and phased DuckHTS paths; it is not an independent VEP oracle.
Both campaigns ran on the Intel i5-13500 with DuckDB 1.5.3, one thread
pinned to CPU 2, three fresh-process timed passes and separate
correctness workers. The current extension was rebuilt from a clean
tracked checkout with HTSlib distclean; its SHA-256 is
a5dceecf83f96b484339fec50372953465dc0543199508bb368df34a9c647f13.
Retained receipts identify source files, native bridge, extension,
registered inputs, derived FASTA/index and every worker artifact. They
are local source-bound records, not signed CI attestations.
Native timing covers initialized ordered-feed replay and a count sink. SQL timing covers full table materialization, including preparation, sort, replay, sequence differences, local coding facts and all nested output. Model/input staging and post-query fingerprints are outside timing. Whole-process maximum RSS includes the R/DuckDB host, setup, warm-up, output and post-query checks.
All measured configurations
All measured medians and process RSS values
| transcripts | samples | overlap | mode | baseline_median_s | baseline_min_s | baseline_max_s | current_median_s | current_min_s | current_max_s | current_max_rss_mib | full_output_equal | replay_equal |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 1024 | 256 | 1 | native | 0.0475 | 0.0457 | 0.0475 | 0.0478 | 0.0467 | 0.0487 | 73.8516 | NA | NA |
| 1024 | 256 | 1 | sql | 1.3070 | 1.3030 | 1.3070 | 1.3110 | 1.3110 | 1.3170 | 957.9570 | TRUE | TRUE |
| 1024 | 4 | 1 | native | 0.0032 | 0.0032 | 0.0032 | 0.0032 | 0.0032 | 0.0034 | 73.8516 | NA | NA |
| 1024 | 4 | 1 | sql | 0.0500 | 0.0490 | 0.0520 | 0.0500 | 0.0500 | 0.0510 | 218.0352 | TRUE | TRUE |
| 1024 | 64 | 1 | native | 0.0138 | 0.0138 | 0.0143 | 0.0135 | 0.0132 | 0.0136 | 73.6953 | NA | NA |
| 1024 | 64 | 1 | sql | 0.3370 | 0.3360 | 0.3410 | 0.3380 | 0.3380 | 0.3500 | 376.6367 | TRUE | TRUE |
| 1024 | 64 | 16 | native | 0.0177 | 0.0177 | 0.0178 | 0.0177 | 0.0171 | 0.0221 | 73.6992 | NA | NA |
| 1024 | 64 | 16 | sql | 0.3430 | 0.3400 | 0.3440 | 0.3480 | 0.3470 | 0.3510 | 378.4023 | TRUE | TRUE |
| 1024 | 64 | 16 | sql_records | 0.7220 | 0.7200 | 0.7260 | 0.7300 | 0.7290 | 0.7370 | 611.9102 | TRUE | TRUE |
| 1024 | 64 | 16 | sql_singletons | 0.3390 | 0.3360 | 0.3420 | 0.3470 | 0.3430 | 0.3570 | 408.9727 | FALSE | TRUE |
| 1024 | 64 | 16 | sql_singletons_hgvs | 0.3480 | 0.3460 | 0.3480 | 0.3490 | 0.3460 | 0.3620 | 409.5078 | FALSE | TRUE |
| 1024 | 64 | 64 | native | 0.0189 | 0.0183 | 0.0190 | 0.0186 | 0.0182 | 0.0190 | 73.8555 | NA | NA |
| 1024 | 64 | 64 | sql | 0.3440 | 0.3430 | 0.3460 | 0.3450 | 0.3440 | 0.3470 | 377.8203 | TRUE | TRUE |
| 10240 | 64 | 16 | native | 0.1713 | 0.1687 | 0.1732 | 0.1705 | 0.1693 | 0.1719 | 73.6992 | NA | NA |
| 10240 | 64 | 16 | sql | 4.0950 | 4.0660 | 4.1270 | 4.1620 | 4.1350 | 4.1630 | 2137.1797 | TRUE | TRUE |
| 10240 | 64 | 16 | sql_singletons | 4.1070 | 4.1070 | 4.1100 | 4.1820 | 4.1650 | 4.1860 | 2190.9141 | FALSE | TRUE |
| 10240 | 64 | 16 | sql_singletons_hgvs | 4.1790 | 4.1700 | 4.1870 | 4.2270 | 4.2160 | 4.2290 | 2191.0586 | FALSE | TRUE |
NA marks the native count sink, which has no serialized SQL-row
fingerprint. Counts, input identities, output contracts, sequence bytes
and JSON byte lengths match between revisions. The unequal singleton
fingerprints mean those timings are same-input, different-output
measurements, not identical-output work. Matching byte lengths do not
establish matching values.
Input and output denominators by measured configuration
| transcripts | samples | overlap | mode | input_records | projected_events | input_calls | input_physical_records | input_record_sample_calls | input_candidate_sample_rows | output_leaves | output_carriers | cds_bytes | protein_bytes | json_bytes |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 1024 | 256 | 1 | native | 4096 | 4096 | 1048576 | NA | NA | NA | 3072 | 458752 | 552960 | 184320 | NA |
| 1024 | 256 | 1 | sql | NA | NA | NA | NA | NA | NA | 3072 | 458752 | 552960 | 184320 | 35028730 |
| 1024 | 4 | 1 | native | 4096 | 4096 | 16384 | NA | NA | NA | 3072 | 7168 | 552960 | 184320 | NA |
| 1024 | 4 | 1 | sql | NA | NA | NA | NA | NA | NA | 3072 | 7168 | 552960 | 184320 | 5142266 |
| 1024 | 64 | 1 | native | 4096 | 4096 | 262144 | NA | NA | NA | 3072 | 114688 | 552960 | 184320 | NA |
| 1024 | 64 | 1 | sql | NA | NA | NA | NA | NA | NA | 3072 | 114688 | 552960 | 184320 | 12185338 |
| 1024 | 64 | 16 | native | 256 | 4096 | 262144 | NA | NA | NA | 3072 | 114688 | 552960 | 184320 | NA |
| 1024 | 64 | 16 | sql | NA | NA | NA | NA | NA | NA | 3072 | 114688 | 552960 | 184320 | 12157566 |
| 1024 | 64 | 16 | sql_records | NA | NA | NA | 256 | 16384 | 262144 | 4096 | 131072 | 737280 | 245760 | 14151208 |
| 1024 | 64 | 16 | sql_singletons | NA | NA | NA | 256 | 16384 | 262144 | 4096 | 65536 | 737280 | 244736 | 9006696 |
| 1024 | 64 | 16 | sql_singletons_hgvs | NA | NA | NA | 256 | 16384 | 262144 | 4096 | 65536 | 737280 | 244736 | 9014888 |
| 1024 | 64 | 64 | native | 64 | 4096 | 262144 | NA | NA | NA | 3072 | 114688 | 552960 | 184320 | NA |
| 1024 | 64 | 64 | sql | NA | NA | NA | NA | NA | NA | 3072 | 114688 | 552960 | 184320 | 12141566 |
| 10240 | 64 | 16 | native | 2560 | 40960 | 2621440 | NA | NA | NA | 30720 | 1146880 | 5529600 | 1843200 | NA |
| 10240 | 64 | 16 | sql | NA | NA | NA | NA | NA | NA | 30720 | 1146880 | 5529600 | 1843200 | 121850254 |
| 10240 | 64 | 16 | sql_singletons | NA | NA | NA | 2560 | 163840 | 2621440 | 40960 | 655360 | 7372800 | 2447360 | 90229832 |
| 10240 | 64 | 16 | sql_singletons_hgvs | NA | NA | NA | 2560 | 163840 | 2621440 | 40960 | 655360 | 7372800 | 2447360 | 90311752 |
Retained output differences
All three measured passes agree within each mode. Across revisions, 12 of 51 timed rows have unequal fingerprints: the two singleton modes at both transcript counts, with three passes each. Literal replay fingerprints agree. The complete observations below retain the unequal hash values; no field was normalized away to obtain a match.
Unequal fingerprint values by configuration
| transcripts | samples | overlap | mode | pass | field | baseline_value | current_value |
|---|---|---|---|---|---|---|---|
| 1024 | 64 | 16 | sql_singletons | 1 | xor_hash | 126084453999582039 | 700836657811329227 |
| 1024 | 64 | 16 | sql_singletons_hgvs | 1 | xor_hash | 2000475952795925397 | 6726181131062349297 |
| 10240 | 64 | 16 | sql_singletons | 1 | xor_hash | 8822573126786218751 | 13918678999014856763 |
| 10240 | 64 | 16 | sql_singletons_hgvs | 1 | xor_hash | 2105622761814596718 | 11764073491056183881 |
| 1024 | 64 | 16 | sql_singletons | 1 | sum_hash | 37582730671888802355371 | 37573251854262826388211 |
| 1024 | 64 | 16 | sql_singletons_hgvs | 1 | sum_hash | 37824893397053031140415 | 37715148776440476629199 |
| 10240 | 64 | 16 | sql_singletons | 1 | sum_hash | 378298692363800329214139 | 377624590454045657951587 |
| 10240 | 64 | 16 | sql_singletons_hgvs | 1 | sum_hash | 378148534262711101761592 | 378079038706658146159825 |
| 1024 | 64 | 16 | sql_singletons | 1 | local_so_xor_hash | 654628605784468392 | 2335123865651189926 |
| 1024 | 64 | 16 | sql_singletons_hgvs | 1 | local_so_xor_hash | 654628605784468392 | 2335123865651189926 |
| 10240 | 64 | 16 | sql_singletons | 1 | local_so_xor_hash | 14224601500847809748 | 5814153375303733475 |
| 10240 | 64 | 16 | sql_singletons_hgvs | 1 | local_so_xor_hash | 14224601500847809748 | 5814153375303733475 |
| 1024 | 64 | 16 | sql_singletons | 1 | local_so_sum_hash | 37446723966248878265056 | 37469739201441226011254 |
| 1024 | 64 | 16 | sql_singletons_hgvs | 1 | local_so_sum_hash | 37446723966248878265056 | 37469739201441226011254 |
| 10240 | 64 | 16 | sql_singletons | 1 | local_so_sum_hash | 377356556873241402705208 | 376754957532195857701877 |
| 10240 | 64 | 16 | sql_singletons_hgvs | 1 | local_so_sum_hash | 377356556873241402705208 | 376754957532195857701877 |
| 1024 | 64 | 16 | sql_singletons | 1 | hgvs_so_xor_hash | 14345163728266764224 | 4389172685361975006 |
| 1024 | 64 | 16 | sql_singletons_hgvs | 1 | hgvs_so_xor_hash | 4662359414007789427 | 2986948762521831407 |
| 10240 | 64 | 16 | sql_singletons | 1 | hgvs_so_xor_hash | 17221334368169699363 | 18088277829535065931 |
| 10240 | 64 | 16 | sql_singletons_hgvs | 1 | hgvs_so_xor_hash | 453159649544162334 | 17894173913969220412 |
| 1024 | 64 | 16 | sql_singletons | 1 | hgvs_so_sum_hash | 37210505766262944422094 | 37319589750075617092114 |
| 1024 | 64 | 16 | sql_singletons_hgvs | 1 | hgvs_so_sum_hash | 37310727840131450206907 | 37380943966290453737923 |
| 10240 | 64 | 16 | sql_singletons | 1 | hgvs_so_sum_hash | 375900478728341038687991 | 376659840003214785142417 |
| 10240 | 64 | 16 | sql_singletons_hgvs | 1 | hgvs_so_sum_hash | 379046938936219164021386 | 379285744703700591171964 |
| 1024 | 64 | 16 | sql_singletons | 1 | non_hgvs_xor_hash | 6052207445973056023 | 18003099339576597538 |
| 1024 | 64 | 16 | sql_singletons_hgvs | 1 | non_hgvs_xor_hash | 6052207445973056023 | 18003099339576597538 |
| 10240 | 64 | 16 | sql_singletons | 1 | non_hgvs_xor_hash | 9551608422294852071 | 2779620574718580229 |
| 10240 | 64 | 16 | sql_singletons_hgvs | 1 | non_hgvs_xor_hash | 9551608422294852071 | 2779620574718580229 |
| 1024 | 64 | 16 | sql_singletons | 1 | non_hgvs_sum_hash | 37933509573515682488189 | 38315864159327482580440 |
| 1024 | 64 | 16 | sql_singletons_hgvs | 1 | non_hgvs_sum_hash | 37933509573515682488189 | 38315864159327482580440 |
| 10240 | 64 | 16 | sql_singletons | 1 | non_hgvs_sum_hash | 377195884730473591625225 | 378704497758867019612841 |
| 10240 | 64 | 16 | sql_singletons_hgvs | 1 | non_hgvs_sum_hash | 377195884730473591625225 | 378704497758867019612841 |
A retained four-event, one-transcript
reproduction isolates
the deletion GC>G at CDS position 67: its block
local_consequence_mask changes from 8,388,608 (frameshift) to
8,396,800 (frameshift plus raw missense). Every other field in those
four complete rows is unchanged. The physical deletion changes local
GCT to partial GT, giving peptide operands A and X. The
documented local-block contract evaluates raw coding predicates without
an independent source event’s uploaded-feature class gates. See
ERRATA.md for the pinned predicate evidence and its
limits. The compact witness preserves both complete native outputs and
all original files from the exact pinned-VEP observation, including
source, observation scripts, commands, receipts and failed-attempt
diagnostics. Its 38 files are stored as byte-exact base64 payloads with
per-file hashes. Rendering verifies the bundle identity, the complete
four-row difference and the independent raw-predicate observation. It is
unsigned diagnostic evidence.
The complete-output audit retains all eight baseline/current Parquet relations for the two singleton modes at both transcript counts. Each untimed capture matches all 27 non-timing result fields from its original correctness worker and all three timed workers, including every full-output fingerprint; Parquet round trips preserve those metrics and types. Rendering verifies the pinned bundle and receipt identities, binds the retained metrics to both ledgers, and checks the exact workload configuration. The original-input capsule preserves 32 original worker jobs, four benchmark receipts, their four FASTA/index pairs, and both registered fixtures. Each original receipt is bound to its capture and ledger; every job’s actual bytes must match that receipt, and the correctness job must also match its capture’s input hash. Decoded jobs must contain the exact registered reference and event table. Their complete semantic fields agree across revisions and passes, with each pass’s correctness flag checked separately; only named binary, output and reference-file paths are excluded from that equality. Extension identities are checked against their own revision, not equated across revisions. The driver remains independently pinned, and retained FASTA/index bytes must match both receipt layers. Original paths are identifiers only: rendering does not reopen historical files. Input controls reject missing, malformed or altered job hashes and self-consistently rehashed jobs with changed events or configuration. Rendering then reruns the complete typed, exact-key comparison.
Complete-output keyed audit counts
| transcripts | mode | joined_rows | missing_keys | changed_rows | changed_other_events | unexplained_rows |
|---|---|---|---|---|---|---|
| 1024 | sql_singletons | 4096 | 0 | 1024 | 0 | 0 |
| 1024 | sql_singletons_hgvs | 4096 | 0 | 1024 | 0 | 0 |
| 10240 | sql_singletons | 40960 | 0 | 10240 | 0 | 0 |
| 10240 | sql_singletons_hgvs | 40960 | 0 | 10240 | 0 | 0 |
Across 90112 keyed row pairs, 22528 differ only in the declared deletion mask above. There are no missing keys, changes to other events or unexplained field changes. The audit checks the full expected transcript/event inventory and deletion geometry, not only aggregate hashes. Its seven corruption controls reject NULL blocks, count-preserving key shifts, altered integer types, changed protein sequence, omitted or duplicate rows and an incorrect deletion mask; rendering reruns them. The bundle retains the observation/comparison scripts, receipts, complete corrupted inputs and original logs. These remain unsigned local observations. The outputs are still unequal: explaining their difference does not turn these measurements into identical-output work or change the historical promotion guard.
Three passes on one shared machine do not establish statistical significance, a causal performance effect or a general absence of regression. This fixture does not measure N-rich annotation, whole-haplotype consequences or compound HGVS.
Reproduction
Use the clean extension-receipt procedure in the baseline report, then
run benchmarks/duckvep_haplotypes.R with --passes 3 --cpu 2 and
these six workloads:
Workload arguments for reproducing each configuration
| transcripts | samples | overlap | modes |
|---|---|---|---|
| 1024 | 256 | 1 | native,sql |
| 1024 | 4 | 1 | native,sql |
| 1024 | 64 | 1 | native,sql |
| 1024 | 64 | 16 | native,sql,sql_records,sql_singletons,sql_singletons_hgvs |
| 1024 | 64 | 64 | native,sql |
| 10240 | 64 | 16 | native,sql,sql_singletons,sql_singletons_hgvs |
Pass --extension-receipt PATH and the listed --transcripts,
--samples, --overlap and --modes to each invocation. Preserve
every result and receipt; do not append a campaign that fails the
historical full-output check.
Canonical workload at the translation checkpoints
Sources 52ce78513b3f5f7c2a46fe82dc44dd458333092b,
de3d008004f615c198a422a42a773bd7c9de0b70,
d6d188da54f0a0cf97ec74244243ebad251194f5,
a4ab9f516275b4c2d1178ddded5da6fbaeccf396 and
6e1a1de9e6cf34dae4dc9f58f27817ca4dd7e8e0 each use the nearest recorded
identical workload as their timing baseline:
20efcf2af33b38c5be7596e2db7b243a3be53c47 for 52ce785, 52ce785 for
de3d008, de3d008 for d6d188d, and d6d188d for a4ab9f5;
6e1a1de compares with a4ab9f5. The workload has 1,024 transcripts,
64 samples, 16 overlapping transcripts, 256 physical events and 262,144
candidate/sample input rows. All six runs use one thread on CPU 2,
DuckDB 1.5.3 and three fresh-process timed passes. No DuckHTS test or
conformance jobs ran concurrently with the 52ce785 timers. No DuckHTS
build, test, conformance or diagnostic jobs overlapped the d6d188d
timers. The same isolation applies to a4ab9f5, measured from 02:53:38
to 02:54:19 +0200 on September 11, 2026. The de3d008 campaign ran from
00:59:19 to 00:59:59 +0200 on September 11, 2026; a small read-only
classification/oracle diagnostic completed at 01:00:01, so concurrent
diagnostic activity cannot be excluded. Other shared-host activity was
not controlled. The 6e1a1de driver ran from 21:28:02 to 21:28:42 +0200
on September 12, 2026, after the build, test and conformance jobs
finished. No other local DuckHTS builds, tests, conformance, benchmarks
or diagnostic computations overlapped its timers; read-only GitHub
status and clock checks continued.
Per-revision translation checkpoint timings
| revision | baseline_revision | mode | input_candidate_sample_rows | output_leaves | output_carriers | baseline_median_s | checkpoint_median_s | checkpoint_max_rss_mib |
|---|---|---|---|---|---|---|---|---|
| 52ce785 | 20efcf2 | native | 262144 | 3072 | 114688 | 0.017662 | 0.017174 | 73.69531 |
| 52ce785 | 20efcf2 | sql | 262144 | 3072 | 114688 | 0.348000 | 0.346000 | 378.93750 |
| 52ce785 | 20efcf2 | sql_records | 262144 | 4096 | 131072 | 0.730000 | 0.731000 | 611.99219 |
| 52ce785 | 20efcf2 | sql_singletons | 262144 | 4096 | 65536 | 0.347000 | 0.343000 | 411.19531 |
| 52ce785 | 20efcf2 | sql_singletons_hgvs | 262144 | 4096 | 65536 | 0.349000 | 0.353000 | 408.82812 |
| de3d008 | 52ce785 | native | 262144 | 3072 | 114688 | 0.017174 | 0.017235 | 73.69922 |
| de3d008 | 52ce785 | sql | 262144 | 3072 | 114688 | 0.346000 | 0.344000 | 379.13672 |
| de3d008 | 52ce785 | sql_records | 262144 | 4096 | 131072 | 0.731000 | 0.734000 | 612.45703 |
| de3d008 | 52ce785 | sql_singletons | 262144 | 4096 | 65536 | 0.343000 | 0.343000 | 409.75391 |
| de3d008 | 52ce785 | sql_singletons_hgvs | 262144 | 4096 | 65536 | 0.353000 | 0.348000 | 409.91797 |
| d6d188d | de3d008 | native | 262144 | 3072 | 114688 | 0.017235 | 0.017650 | 73.85156 |
| d6d188d | de3d008 | sql | 262144 | 3072 | 114688 | 0.344000 | 0.347000 | 378.92188 |
| d6d188d | de3d008 | sql_records | 262144 | 4096 | 131072 | 0.734000 | 0.738000 | 611.44922 |
| d6d188d | de3d008 | sql_singletons | 262144 | 4096 | 65536 | 0.343000 | 0.342000 | 411.05078 |
| d6d188d | de3d008 | sql_singletons_hgvs | 262144 | 4096 | 65536 | 0.348000 | 0.348000 | 410.01953 |
| a4ab9f5 | d6d188d | native | 262144 | 3072 | 114688 | 0.017650 | 0.016827 | 73.85547 |
| a4ab9f5 | d6d188d | sql | 262144 | 3072 | 114688 | 0.347000 | 0.346000 | 378.33203 |
| a4ab9f5 | d6d188d | sql_records | 262144 | 4096 | 131072 | 0.738000 | 0.728000 | 611.23828 |
| a4ab9f5 | d6d188d | sql_singletons | 262144 | 4096 | 65536 | 0.342000 | 0.344000 | 408.10938 |
| a4ab9f5 | d6d188d | sql_singletons_hgvs | 262144 | 4096 | 65536 | 0.348000 | 0.348000 | 408.94922 |
| 6e1a1de | a4ab9f5 | native | 262144 | 3072 | 114688 | 0.016827 | 0.017450 | 73.69922 |
| 6e1a1de | a4ab9f5 | sql | 262144 | 3072 | 114688 | 0.346000 | 0.344000 | 378.75000 |
| 6e1a1de | a4ab9f5 | sql_records | 262144 | 4096 | 131072 | 0.728000 | 0.734000 | 612.30859 |
| 6e1a1de | a4ab9f5 | sql_singletons | 262144 | 4096 | 65536 | 0.344000 | 0.340000 | 409.71094 |
| 6e1a1de | a4ab9f5 | sql_singletons_hgvs | 262144 | 4096 | 65536 | 0.348000 | 0.348000 | 409.24219 |
Each checkpoint’s 15 matched rows have identical non-timing result fields, including every SQL full-output fingerprint, output denominator and native workspace/count metric. The 75 checkpoint observations retain individual passes. The 525-file evidence capsule retains all six runs’ original worker jobs, results, logs, process-time reports, benchmark receipts and FASTA/index pairs, all six clean-build receipts, the registered fixtures and the unchanged driver. Rendering verifies every retained file against its receipt before comparing actual jobs and results. It requires complete job semantics to agree across revisions, excluding only the explicitly checked correctness flag and named binary/output/reference-file paths. All 120 result objects are checked for within-mode repeatability; all timed values and RSS measurements are bound to the corresponding ledger rows. Sixty job-mutation controls reject changed sample counts or event alleles. Original paths remain identifiers only; rendering does not reopen historical files.
Compiled extension and native-bridge payloads are not included in this
capsule; their identities are retained and checked through the original
receipts. The 6e1a1de extension SHA-256 is
a53dcb1e1046e8bcb414400c69201d428e96a56a763dfee531fee7dc399bca9d.
These are unsigned local source-bound measurements. The historical
unequal-output comparison and its complete-output controls remain
separate and unchanged. The checkpoint comparisons establish fingerprint
equality, not complete typed-row equality.
This canonical table-1 workload measures shared replay and
singleton-HGVS paths. It does not measure N-bearing REF, skipped raw ALT
or nonstandard-codon-table branches, including the N/U/lowercase
REF-slot correction in 6e1a1de. Three passes on a shared host do not
establish statistical significance, causality or a general absence of
regression. Reproduce with the clean-build receipt procedure above and
this command:
Rscript benchmarks/duckvep_haplotypes.R \
--transcripts 1024 --samples 64 --overlap 16 --passes 3 --cpu 2 \
--modes native,sql,sql_records,sql_singletons,sql_singletons_hgvs \
--extension-receipt /tmp/duckhts-foundation-merge.a0qkeM/extension.tsv