question:  Does any arm, alone or in a portfolio, reach instances the shipping
           otf_sparse_formula configuration does not -- and how much of what the
           arms reach alone survives racing them under one fixed budget?
           Written up in benchmarking/ADAPTIVE-PORTFOLIO-OTFUR-SPRINT.md, sections 6-9.

acacia:    b6da274fa4468cf6b1dd82bd3ae0d8aa8cdb5fb2 (the sprint's patch stack on 3fb9f113)
           preset otf_sparse_formula, self-benchmark.sh -R (release profile: -O3, LTO,
           acacia_compiler_profile=release, so -Ofast -march=native -- not portable)
           binary sha256 5451c91e2a133ac5a51ceb7294d3892e923a58727a2a7e49477e1376ca4c5f73
baseline:  3fb9f113d3c47294ae3527092850b90f59ec0635, same preset and profile
           binary sha256 2d5707332ec4ff2e1347598825f84e6ef56c350a1c52b4edc50a76496719ade4
           used for the infinite-race-u10 A/B (report section 8); kept built for G3
           NOTE: rebuilt 2026-09-12 03:04 while fixing G1, so the binary now in
           acacia-prepatch hashes d9db7dab83e1d8ab..., not the sha256 above. Same
           revision and preset, but the release profile is not reproducible. The
           candidate gate binary ba3159daef74bd1e... was removed with its worktree.
posets:    139e143
spot:      2.15.1.dev

harness:   a pinned copy of benchmarking/run-syntcomp26-coverage.py and benchlib.py taken
           from 00a066ae, run from outside the work tree so that a branch switch could not
           change it mid-campaign
             run-syntcomp26-coverage.py  94ae42592693d98094c664613722bb8e98c5b14d4f405d004852fdee80350ecf
             benchlib.py                 999a35829a81f4a5176515ec27253dbfa77188cf0e2f3e4f128db4053e9ca4fb
           passed explicitly, because from a pinned copy both would resolve wrongly:
             --status-exceptions benchmarking/syntcomp26-status-exceptions.tsv
             --acacia-sha b6da274fa4468cf6b1dd82bd3ae0d8aa8cdb5fb2
           every arm ran with --candidate-mode only

corpus:    tests/suites/benchmarks/syntcomp26/tlsf-sources.tsv over the materialized flat
           TLSF corpus (1,586 files)

host:      11th Gen Core i7-11850H, 8 cores / 16 threads, 15 GB, zram swap ~7 GB in use.
           Little headroom above the 8 GiB cap: trust MEMOUT rows less than timeouts.

panel-census/   22 arms (arms.list), each alone, over
                tests/suites/benchmarks/syntcomp26/panel.list (180 instances).
                Staged caps 1,5,17; read at cap <= 17 through smallest_cap_solved.
                8 GiB MemoryMax, zero swap, one invocation at a time, no CPU limit.
                2026-09-10 13:07 to 2026-09-11 01:32.

hard-set/       10 arms (arms.list), each alone, over instances.list: the 353 of 1,524
                that otf_sparse_formula leaves undecided in the closing threeway campaign,
                _bm-logs.spot-otf-20260907/full-closing/otf_sparse_formula-summary.tsv
                (1,171 decided at a uniform 17 s).
                Uniform 17 s cap, 8 GiB, zero swap, one invocation at a time, no CPU limit.
                2026-09-11 01:33 to 19:12. Arm 4 (unreal:formula:backward) was interrupted
                by a session restart after 22 instances and resumed with --resume; rows are
                fsynced per instance, so the one in flight was simply re-run.

race-curve/     One race invocation per instance per config (configs.list) over
                instances.list: the panel plus the hard set, 494 after an overlap of 39.
                Uniform 17 s, 8 GiB for the whole race, zero swap, CPUQuota=400%.
                AllowedCPUs was tried first and is silently ignored here, because the user
                manager is not delegated cpuset; that attempt ran a few minutes unpinned and
                is kept as _bm-logs.20260911-race-curve-UNPINNED-partial, not evidence.
                Summaries are added here when the run completes.

not here:  the raw per-run TSVs, which stay in the gitignored _bm-logs.20260910-arm-census,
           _bm-logs.20260910-unlocked-coverage and _bm-logs.20260911-race-curve.
