Drive an archive from a printout, and check it on a scanner

[bbfrederick]

Three commits taking the exar driver from "produces a file" to "produces a file a scanner accepts", and fixing what the scanner's return exposed.

The driver's output has now been loaded

examples/XA60/driver_loadtest.{exar1,pdf} is Potpourri_P1 driven by the Potpourri_P1_changed printout, loaded on an XA60 scanner and re-exported. 18 scans in, 18 out, running order preserved, program still named DRIVER_TEST, nothing greyed out.

That is what the answer-key comparison could not establish. Only ~30% of what a protocol prints has a verified mapping, so the result is a hybrid — 226 fields from the printout beside 173 inherited from the template — and a hybrid of two consistent parameter sets is not itself consistent. For this protocol it is.

ASCCONV fields the driver wrote	226
...returned byte-identical	226
...that genuinely moved off the template	226
Printed parameters the console changed	87
...reproduced in our build	80 (92%)

insert_ascconv is validated on a scanner for the first time: sWipMemBlock.adFree[3] is absent from the template, was created between its siblings, and came back in place with the console's spelling.

And it exposed a defect that field-by-field agreement could not

Slice Thickness went 2.3 → 2.2 across all 64 elements of a CMRR EPI while sPosition kept describing the 2.3 mm geometry — the outermost slices (64-1)/2 × 0.1 = 3.15 mm from where they belong. The console recomputes all 64 on its own edit; we did not.

The comparison above agreed on all 226 written fields and could never have found this, because the defect is in fields nobody wrote: correct in the template, correct in what we sent, correct in what came back, and wrong relative to the thickness beside them. The corpus sweep found it, which is the argument for shipping the return as an example.

The scanner did not catch it either. The scan loaded, was not greyed out, printed no complaint, and returned the array untouched. So the console's consistency check does not cover the slice array against its group inputs — Include Nav. = Off gets refused, an incoherent 64-slice group does not. "It loaded" is a weaker signal than the greying-out rule suggests.

build.recentre rebuilds an array this write invalidated, leaves one that arrived broken alone, and skips multi-group arrays. Against the same answer key: 201 fields exact, up from 158.

The return keeps its bad array, pinned as KNOWN_INCONSISTENT_ARRAYS rather than excluded, so a second one fails.

Reading the direction letter

A Siemens printout gives a coordinate as a magnitude beside a direction letter — F32 for a stored -32 — and on a two-column card the letter lands in a field of its own. Table Position was reading the number alone, which flipped the sign on every scan whose position is negative. Mapping.sign_from now reads the letter: H against non-negative, F against negative, on all 300 corpus comparisons, and a letter that is neither is refused rather than defaulted.

Also fixed: expand searched for the mapped field across array indices instead of asking which elements exist, so a sparse member like dDistFact was refused with a reason that was not true; and lRepetitions was not marked sparse, making Measurements unwritable on 585 scans. Both surfaced from the manifest reporting refusals.

Coverage the pair adds

With its printout beside it, driver_loadtest joins two corpus sweeps:

• Self-drive — 18 scans matched, 0 values written. First time that invariant runs against a protocol this package wrote rather than one it only read.
• Flag bits — 50 comparisons against a card the scanner printed from what we stored. Every prior comparison decoded a console's word.

One actionable gap named

Averaging is mapped for tfl_mgh_multiecho and printed by tfl_mgh_epinav_ABCD, so the driver refuses it — correctly, since a sWipMemBlock index means whatever its own sequence says. Decoding alFree[4] with the same table nonetheless agrees on 37 of 37 corpus scans across both sequences. That is evidence of the wrong kind: printout-against-stored tests the decoder, since the console displays what it stores. Settling it needs an option scan varying Averaging on the second sequence.

BuildReport.out_of_scope now names that class instead of burying it among the forty parameters nothing has looked at. build.covered_elsewhere asks the mapping table, not resolve's prose, so rewording a reason cannot reclassify anything.

Verification

• 2486 passed, 19 skipped on the full suite
• black, isort, codespell clean
• New tests: driver round-trip (writes survived; scanner changed nothing unexplained), driver geometry consistency with a mutation check, recentre's three branches, and the manifest classification

🤖 Generated with Claude Code