479. The symbol level

2026-09-09 - directed, continuing 478

Built the second half of the differential: symbol counts, import triples, relocation census.

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

What it compares

• The dynamic symbol count, from two independent derivations - orbistoun from DT_HASH's nchain, SELFish from symtabsz / syment. Agreeing is only worth something because neither read the other's number.
• Every encoded import, joined on symbol index, comparing hash and resolved library and module names. The index is the join because a relocation names its symbol by index and nothing else.
• The relocation census by type, for rela and jmprel.

All three agree, on all 29 modules.

Three meanings read out of the source first

478's lesson applied rather than repeated: SELFish's imports returns encoded names only, so plain undefined symbols are counted apart; SELFish resolves library and module to names where orbistoun carries raw ids, so the names are what gets compared; and the NID byte order.

The byte order is measured, not assumed, and it had to be: SELFish's answer to REQ-20260909T1250Z-1f74 says orbistoun prints these reversed, while orbistoun's own NidHasher::hash documents a little-endian read that sounds like agreement. Two documents, opposite implications. The test tries both and counts which matched: every encoded import in every module matches only reversed, no module mixing the two. A module that mixed them is a separate reported field, not a silent majority vote (D655).

Surprises

• Reporting the units as disagreements said the opposite of the truth. 28 modules "differing" when they agree about all of them. Disagreement now carries a units flag and the headline counts them apart - the line still prints, because a convention nobody wrote down is how three of 478's four errors happened.
• Two vacuous passes were possible and both are now loud. symbol_count returns None when symtabsz is zero, which would skip the comparison; an empty relocation census compares equal to an empty one. Neither fired on this corpus. Both were written before knowing that.
• The doc-comment mistake again - inserted a function above an existing one and split its doc comment off its item. Fourth time. Insert after the closing brace, never before a comment.

Next

• Reported back to SELFish as REQ-20260909T1910Z-3d05, asking them to write the byte-order convention down on their side.
• The exception context at +0xf8, still open, still the wall for the furthest-reaching title.