D355 - A turn that measured a contract and wrote nothing

decided · 2026-08-27 · asked why three diagnosed titles produced no patches

Three titles were turned in one sitting. One found a real contract - sceLibcMspaceMalloc answering the code the guest followed, zero reaching 25 against 13 - and nothing was written. The finding existed as terminal output, which CLAUDE.md names as already lost.

The cause is a caution applied to the wrong act. Two flags:

flag	what its own doc says
--record	"deciding to change a tracked file stays a deliberate act with a diff"
--apply	"Not a tracked file... deleting it is a complete undo"

--apply is gated by the tracked-file caution, and what it writes is not tracked. So a turn given neither flag - the ordinary case - persisted nothing at all.

Emitting and applying are different acts and only one needed gating. A proposal is a file nothing applies, in an untracked directory, undone by deleting it. Applying changes what the next run does, which is the act that needs an oracle behind it. turn now always writes what it measured to patches/, and --apply still gates the policy change (D322).

What had to be fixed to make it possible

unimplemented_calls stripped library:: from every candidate, because ORBISTOUN_RETURN splits its value on : and cannot express a qualified name. True of the variable, and it threw the library away for everything downstream - so the measurement could not say which knowledge file it belonged in and the proposal was skipped. Stripped where the axis is built now, which is the only place that needs it.

The bug the fix uncovered, which was already there

Adding recording to turn filed this:

[experiment]  outcome = "ran to the time limit"  imports = 25  standing = 55

For a title that reaches 13. The number was bought by a reserved region the guest never asked for - a diagnostic run, recorded as a compatibility claim. D227 says it directly: an intervention that moves a wall is not a diagnosis.

record_compat now refuses an intervened run. The hazard was already there for run - anything with a diagnostic set would have filed the same kind of number - and nothing had exercised it until a turn did.

And the gap I diagnosed did not exist

turn was said to be missing the recording run does. It is not: GuestTrial shells out to orbistoun-cli run for every boot, so each boot already records itself - the baseline honestly, and the intervened ones now refused. The record_compat added to cmd_turn was redundant, read the last trace (a diagnostic), and printed "not recorded" on every turn while the baseline had recorded fine. Removed.

Worth keeping because the investigation was worth more than the diagnosis: a claimed gap that was not real led to a real bug two layers away, and the way it surfaced was the guard printing "not recorded" while the file changed anyway.