P-M0-2: the gitingest salvage rate holds at 6/8 or better

id:        P-M0-2
filed:     2026-07-17, migrated from the journal-format filing in the M0 milestone note;
           the settling run has not launched
suite:     swebench-live, cyclotruc__gitingest-94, eight single-attempt repeats
tier:      free (deepseek-v4-flash-free)
engine:    tomo-oi at the M0-landed pin
band:      floor: 6 of 8 passes.
           ceiling: the journal band of 7/8, give or take one task of sampling slack.
mechanism: the salvage layer recovers real commands the cheap model emits in
           non-markdown fences (the four observed costumes), which is what took this
           task from about 2/8 to 7/8; that machinery is untouched by M0's other
           changes, so the rate should hold.
settled by: eight single runs at the landed pin.
           go run ./cmd/lab run tomo-oi cyclotruc__gitingest-94 --suite swebench-live
result:    (open, blocked on free-tier quota reset)
verdict:   (open)

This entry predates the ledger machinery and migrated in from the M0 milestone note, filed before its settling run, which has still not launched.

The prediction deliberately gives back one task of slack against the journal's 7/8 rather than pinning the exact number. Below 6/8, the salvage slice is re-examined before M0 closes.