Bump reconverse to 58921e9 (#217, #222); grow the reconverse CI tier; skip classic-only tests on reconverse

[lvkale]

Summary

Bumps contrib/reconverse to main 58921e9, which carries two fixes found by Charm++'s own tests, and lets CI run those tests.

• reconverse Dummy issue #217: CmiWithinNodeBroadcast follows the classic delivery contract (nokeep shared by reference, zerocopy broadcast receive delivered to the calling PE only, caller's buffer consumed). Before it tests/charm++/within_node_bcast aborted at +pe 2 and six of nine tests/charm++/zerocopy directories segfaulted at +pe 4.
• reconverse Port miniFE to Charm++ #222: reconverse declares CMK_ONESIDED_IMPL (with CMK_ONESIDED_RO_DISABLE), so zcQdIncrement creates one quiescence count per Direct-API RDMA operation, matching reconverse's single acknowledgement. Before it, zerocopy_with_qd with two processes never reached quiescence after its Direct API test. It also declares CMK_MACHINE_PROGRESS_DEFINED, so CkNetworkProgress() drives the network instead of compiling to nothing.

CI tier (reconverse-ci.yaml TEST_DIRS): adds within_node_bcast and eight zerocopy directories, the tests that caught both bugs. zerocopy/zc_post_modify_size stays out until reconverse #223 (two-process data mismatch, present before and after these fixes) is resolved.

Classic-only tests skipped on reconverse (CMK_GDIR=reconverse): tests/charm++/queue (classic's Cqs API, not compiled for reconverse), tests/charm++/longIdle (needs CcdCallOnConditionOnPE), tests/converse/megacon (Cpm interface; reconverse carries its own port, tests/megarecon, reconverse PR #218).

No Charm++ source changes; the quiescence fix takes effect through the rebuild.

Verification

macOS arm64, this tree: the whole TEST_DIRS tier passes in one and two processes, zerocopy_with_qd included. An Anvil run of the same reconverse (two nodes, InfiniBand, hwloc) is in progress and should gate the merge; results will be posted here.

🤖 Generated with Claude Code

[lvkale]

Anvil results (one node so far; two-node jobs queued behind ~23k pending): charm 310d470e0 built against reconverse 75703b2 (main + #217 + #218 + #220 + #222), EPYC 7763 node, LCI ibv, hwloc 1.11.13.

• Charm++ tier: 19/19 programs pass single-process, 18/19 with two processes. zerocopy_with_qd reaches all six QD phases cross-process (the Port miniFE to Charm++ #222 fix). The one failure is zc_post_modify_size with two processes: Network completion failed then a hang, i.e. reconverse Replica on Titan is slow #223, present before and after these changes and excluded from the tier here.
• Affinity binds correctly with real hwloc (+setcpuaffinity +pemap, +showcpuaffinity prints the PU per PE).
• Reconverse ctest on the compute node: 54/56, the two failures being a test-definition bug already fixed in reconverse Dummy issue #220 and an oversubscription timeout in an 8-CPU allocation (passes with 16).
• New, pre-existing: anytime_migration/migrate aborts at +pe 3 and +pe 4 in one process (CmiPushPE(... destPe: -1 ...): rank out of range), on the Mac too; +pe 2, which the Makefile runs, passes. Being written up separately.

Two-node results (InfiniBand across hosts) will be added when the wholenode jobs run.

[lvkale]

We don't need to wait for 2 node result for this. Please review and merge.

[lvkale]

Anvil two-node results are in (wholenode a[365,375] and standard a[492,510], InfiniBand between hosts, charm 310d470 with reconverse 75703b2 = main + #217 + #218 + #220 + #222): the whole tier passes across two nodes, including zerocopy_with_qd (all six QD phases) and within_node_bcast. The only nonzero results are the known ones: zc_post_modify_size across nodes times out (reconverse #223, present before these changes); anytime_migration at +pe 4, which its Makefile does not run (the test assumes 2 PEs); and +pemap 0-3 on the standard partition, whose cgroup excludes cores 0-3 (passes on wholenode). Reconverse's own ctest across two nodes: 35/35 single-process, 19/21 multi-process, the two being the abort_peer wrapper already fixed in reconverse #220 and a partitions-two timeout (reconverse #213's test; passed on one node, to be looked at separately).