cefacd2d0 (#641, the LTX-2.5 landing) left
tests/scripts/test_agent_record.py::test_model_row_ratchet_is_load_bearing
with two contradictory paragraphs spliced into one docstring. The diff is
purely additive — nothing was replaced, four lines were appended:
$ git show cefacd2d0 -- tests/scripts/test_agent_record.py
seven recipe architectures that had no row at all took it 362 -> 369
(#609, #610). Without this, bumping the number to silence a failure is
indistinguishable from bumping it because a row really landed.
+ architecture re-pins it by hand. Muse Glimmer took it 361 -> 362, and
+ LTX-2.5 took it 362 -> 363. Without
+ this, bumping the number to silence a failure is indistinguishable from
+ bumping it because a row really landed.
The appended half restates a truncated sentence and then states a transition
that never happened. LTX-2.5 did not take the MODEL pin 362 -> 363. From
git log -S on scripts/check-agent-record.py:
| Commit |
Change |
Taken for |
c8fc24a50 |
361 -> 362 |
Muse Glimmer |
eba6ab7c7 |
362 -> 369 |
the seven recipe architectures (#609, #610, PR #625) |
cefacd2d0 |
369 -> 370 |
LTX-2.5 (#435, PR #641) |
The pin is 370 today (scripts/check-agent-record.py:68), so 362 -> 363
contradicts the tree as well as the history.
Why it matters beyond tidiness. This docstring is the only record of what
each bump was taken for, which is the entire point of the ratchet: the test
exists so that "the number moved because a row landed" cannot be confused with
"the number moved to silence a failure". A wrong provenance line makes the next
reader's audit start from a false premise.
Cause is a merge resolution: the LTX-2.5 branch predated #625, where the pin
really did go 362 -> 363 on that branch, and the stale paragraph was appended
rather than dropped when main's 362 -> 369 arrived. No checker looks at docstring
prose, which is why it landed green.
Repair: one coherent paragraph naming all three transitions with the commit that
made each. Records/test-prose only; no assertion changes.
cefacd2d0(#641, the LTX-2.5 landing) lefttests/scripts/test_agent_record.py::test_model_row_ratchet_is_load_bearingwith two contradictory paragraphs spliced into one docstring. The diff is
purely additive — nothing was replaced, four lines were appended:
The appended half restates a truncated sentence and then states a transition
that never happened. LTX-2.5 did not take the MODEL pin 362 -> 363. From
git log -Sonscripts/check-agent-record.py:c8fc24a50eba6ab7c7cefacd2d0The pin is
370today (scripts/check-agent-record.py:68), so362 -> 363contradicts the tree as well as the history.
Why it matters beyond tidiness. This docstring is the only record of what
each bump was taken for, which is the entire point of the ratchet: the test
exists so that "the number moved because a row landed" cannot be confused with
"the number moved to silence a failure". A wrong provenance line makes the next
reader's audit start from a false premise.
Cause is a merge resolution: the LTX-2.5 branch predated #625, where the pin
really did go 362 -> 363 on that branch, and the stale paragraph was appended
rather than dropped when main's 362 -> 369 arrived. No checker looks at docstring
prose, which is why it landed green.
Repair: one coherent paragraph naming all three transitions with the commit that
made each. Records/test-prose only; no assertion changes.