The module docstring claimed "every value asserted here is a named row of a
named table", but the converse did not hold: 27 of the 83 shipped rows had a
dimension no test compared against a literal. ISO 4014/4017 M5 (s,k), ISO 4032
M5 (m), ISO 7089 M5 (d1,d2,h), ISO 4762 M4/M6/M10/M16 (dk,s,b) and D/B for 20
of the 33 ISO 15 bearings were constrained only by inequality tests (series
growth) and by the bore-number rule - never by the transcribed number. For a
task whose worst failure is "a standards table row that is not in the
standard", those rows were exactly where such a row would hide.
Measured before the fix with a pytest plugin that rewrote five rows in memory
(M5 hex bolt s=k=99, M5 nut, M5 washer, M16 socket screw dk=s=b=99, 6307 as
78x20 instead of 80x21): 32 passed in 0.07s. All five now fail.
Each of the five tables is pinned by an ISO*_ROWS dict literal carrying every
dimension of every shipped size, parametrized one case per size/designation,
plus a completeness test asserting the table's key set equals the literal's -
so a row added later without a pin is red too (verified: a fabricated M14 head
and 6011 bearing fail the completeness gate). 32 -> 93 tests, 0.06 -> 0.13 s.
No production code changed; no ActiveX member is involved, so no live probe was
required for this finding.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FKFStjEFzpGyLq5bMpWxTd
