# The small-instance view: the instances the M5 criterion "small integer instances are
# solved to their published optimum" is about, with a node budget large enough for a
# proof to be possible.
#
# The list is the manifest's instances with at most 300 rows, minus the ones whose
# measured cost per node makes a 30 000-node run unaffordable. The costs below were
# measured on this machine with the native release build, root cuts on, 300 to 1 000
# nodes per instance:
#
#   flugpl      18 vars   18 rows   0.4 ms/node  (proves optimality in 13 806 relaxations)
#   gt2        188 vars   29 rows   1.6 ms/node
#   noswot     128 vars  182 rows   3.1 ms/node
#   blend2     353 vars  274 rows   7.2 ms/node  (proves optimality in 24 934 relaxations)
#   p0201      201 vars  133 rows  32.4 ms/node  (proves optimality in 609 nodes)
#   khb05250  1350 vars  101 rows  18.3 ms/node  (proves optimality in 107 nodes)
#   22433      429 vars  198 rows   (proves optimality in 33 nodes)
#   markshare1  62 vars    6 rows
#   markshare2  74 vars    7 rows
#   pk1         86 vars   45 rows
#
# Excluded, and why (the arm's cost is the measured per-node cost times the budget):
#   dcmulti    548 vars  290 rows  87 ms/node  -> a 30 000-node arm is ~44 minutes
#   rout       556 vars  291 rows 391 ms/node  -> a 30 000-node arm is ~3.3 hours
#   misc07     260 vars  212 rows over 8 minutes for 1 000 nodes -> over 4 hours
#   50v-10    2013 vars  233 rows  (over the variable ceiling of this view)
#   mod010    2655 vars  146 rows  (over the variable ceiling of this view)
#
# Nothing is hidden by leaving them out: bench/mip-report.md covers every instance in
# the full manifest at a 300-node budget, and reports these as node limits there.
#
# blend2 was excluded for fifteen rounds because it was blocked rather than expensive:
# its no-cuts run stopped at node 30782 on a certificate the verifier refused, first on
# the dual signs. That is fixed in the kernel now (the stop decision measures the
# multiplier sign convention the certificate is read in, and a basis that cannot be
# written down is reported as a numerical failure instead of being emitted), so it is
# back in this view; see CHANGELOG.md and docs/roadmap.md.
flugpl
gt2
blend2
khb05250
markshare1
markshare2
noswot
p0201
pk1
22433
