458. The ratchet cannot say worse

2026-09-08 - directed, continuing 457

Bus idle a seventh pass. Two decisions still with the operator. So: verify the corpus, which turned out to be the productive thing to do.

PPSA28061 has lost 44% of its imports and nothing said so

Running the whole corpus to confirm nothing regressed found something that had:

PPSA28061-app0   26 distinct, 334 calls, the guest called abort
committed record  47 imports, 933 calls, image+0x43c4, measured_on 2026-08-23

Stable across three runs. Not this session: git stash -u, run on a clean HEAD, restore - same 26, same abort. Sixteen days old at least.

Neither instrument can express it (D634). The compat record is a ratchet - beats gates every write, for the good reason that a looser stub policy would otherwise overwrite an honest entry permanently - so a worse run leaves no trace. And the run verdict compares against the previous run, so once a regression is the new normal every run afterwards says (+0) and same, forever.

Between them they answer "did this beat the last run" and "what is the best ever seen", and neither answers "is this build below its own record".

The frontier's own test uses status(Reach::Entered, 47, 933) as its example of an honest record. That is this title's number, and the build that produced it is gone.

Three walls, one subsystem

Consolidated separately (D633), because three findings from three directions turned out to be one decision rather than three: a title's own mapped code answered with a placeholder (D630), a symbol that should have been bound left null (D628), and - established this pass - a symbol that should be null, bound.

That last is 900-surface/control: obSCEne checks its own instruments with a weak symbol it defines and one it deliberately never defines. The console reports the second absent. Orbistoun reports both present, so the probe declares its own census meaningless - which is why ninety-four census entries carry a warning rather than a reading. Neither symbol is an import; the mechanism is relocation-shaped and which relocation has not been established, so it is written at that strength.

All three need a change to import resolution, which this session's brief places outside the work.

Surprises

• The corpus check was the finding. It was meant to confirm nothing broke.
• PPSA28061's [experiment] record reached 60 imports under a diagnostic, against 47 honest and 26 today. So the title has been measured three ways and the two useful numbers are both in the past.

Next

• Give report_progress the committed record, so a run below its own best says so. This would have caught the above the first time it happened.
• Bisect PPSA28061 against the build of 2026-08-23 - the record carries the date, so the method is clear.
• The three relocation walls (D633), waiting on scope.
• ORBISTOUN_DLSYM_STUBS as a default (D632), waiting on a decision.