D655 - The symbol level agrees

Status: measured Date: 2026-09-09

What was compared

D654 left the table level clean and the symbol level unreachable. With every module now getting past the tables, the differential compares three more things per module:

• The dynamic symbol count, from each reader's own derivation. Orbistoun takes it from DT_HASH's nchain; SELFish divides symtabsz by syment. Two independent readings of one file, which is the only reason comparing a count is worth anything.
• Every encoded import, joined on symbol index, comparing the hash and the resolved library and module names. The index is the join because a relocation names its symbol by index and by nothing else.
• The relocation census by type, for rela and jmprel.

All three agree on all 29 modules.

differential: 29 module(s), 0 with a disagreement, 28 differing only in units

Three meanings established before a line of comparison was written

D654 cost four rounds to one mistake - two fields sharing a name and not a meaning - so each of these was read out of the two crates' source first:

• SELFish's imports returns encoded names only. A plain undefined symbol, what an open toolchain emits, is skipped there and carried here as NameForm::Plain. Counting both lists would report a difference on every module that has one.
• The library and module come out differently: orbistoun carries raw ids and resolves them through its own tables, SELFish resolves them for you. The names are the comparable thing, and the one worth comparing - a wrong id attributes an import to the wrong module and produces a name that fits and means nothing, which is D117 and cost a week.
• The NID is byte-reversed between the two.

The byte order is measured, not assumed

This is the one that would have been assumed a fortnight ago. SELFish's answer to REQ-20260909T1250Z-1f74 said orbistoun prints these reversed; orbistoun's own NidHasher::hash documents "the first eight bytes of the SHA-1 digest, read little-endian", which reads as though the two should agree. Two documents, opposite implications, and picking one would have been the D654 error again.

So the test tries both orders, counts which matched, and reports it. Every encoded import in every module matches only reversed - 583 in PPSA02664's eboot, 1,733 in PPSA21564's - and no module mixes the two. A module that did mix them would mean neither reader has a convention, and that is a separate reported field rather than a silent majority vote.

Units are counted apart from disagreements

Every module with encoded imports differs on byte order. Reported as a disagreement that reads as the two readers disagree about 28 modules, when they agree about all of them. Disagreement carries a units flag and the headline counts them separately - the line still prints, because a convention nobody wrote down is how three of D654's four errors happened.

Two vacuous passes closed

A comparison that never ran and a comparison of nothing both look exactly like agreement, which is the failure this whole test exists to catch one level down. Two were possible here and both now report themselves:

• symbol_count answers None when symtabsz or syment is zero, which would skip the comparison silently. It answered on all 29, and the skip would have said so.
• An empty relocation census compares equal to an empty one. Neither table was empty anywhere, and a pair that was would now print rather than pass.

Both were written before knowing they would not fire. That is the point of them.