You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
gc-ratchet baseline has mixed provenance after #7650's surgical re-pin, and the schema cannot express it #7652
benchmarks/gc_ratchet/baseline/gc-ratchet-v1.json now has mixed provenance: 143 of its 144 cells were measured at 26b9c9d59 (0.5.1346), and 01_nursery_churn.heap_used_bytes was surgically re-pinned at c8394bfdb+#7650 (0.5.1370). Same host both times, verified against the artifact's own host block.
That was the right call for #7650 — the alternative was spending probe_overrides' one-way gating: false on a deterministic cell, whose premise (genuine sample-dependence, per 12_large_live_set's rationale) does not hold here. And it is recorded: the artifact's notes says which cell came from where, and the top-level commit is described as covering 143 of 144.
The problem is what a future reader assumes. A baseline artifact carrying a single top-level commit field reads as uniform provenance, and the note that says otherwise is 4,000 characters into a prose block. The next person to attribute a moved cell will reason against the wrong commit for one row in twelve probes.
Make mixed provenance structurally visible rather than prose-visible — a per-cell or per-probe commit field that validate understands, so a surgical re-pin records its own provenance where the reader is already looking. Today the schema cannot express what was actually done.
Until (1), treat 26b9c9d59 as the provenance for every cell except01_nursery_churn.heap_used_bytes.
Closed by #7657. The surgical re-pin's mixed provenance is gone: #7558's fix drove every probe's conservative-scan excess to 0.00%, so the whole artifact was re-pinned on the pinned mini in one pass and probe_overrides is now {}.
Worth recording what that means beyond provenance: 12_large_live_set.heap_used_bytes carried a one-waygating: false spent specifically on this residue. It is gated again. A gating exemption is designed never to come back, so reclaiming one is the strongest possible evidence the underlying cause was real and is now gone — rather than tolerated.
The schema request in point 2 (make mixed provenance structurally visible rather than prose-visible) no longer has a live instance, so I am not keeping it open speculatively. If a future surgical re-pin is ever the right call again, this issue is the argument for adding a per-cell commit field at that time.
Activity
proggeramlug commentedon Aug 8, 2026
Closed by #7657. The surgical re-pin's mixed provenance is gone: #7558's fix drove every probe's conservative-scan excess to 0.00%, so the whole artifact was re-pinned on the pinned mini in one pass and
probe_overridesis now{}.Worth recording what that means beyond provenance:
12_large_live_set.heap_used_bytescarried a one-waygating: falsespent specifically on this residue. It is gated again. Agatingexemption is designed never to come back, so reclaiming one is the strongest possible evidence the underlying cause was real and is now gone — rather than tolerated.The schema request in point 2 (make mixed provenance structurally visible rather than prose-visible) no longer has a live instance, so I am not keeping it open speculatively. If a future surgical re-pin is ever the right call again, this issue is the argument for adding a per-cell
commitfield at that time.