Remote compute-node workflow

[BenWibking]

A remote server currently starts each dataset's block cache with the client's budget (normally 1 GiB), regardless of the memory allocated to its Slurm job. This adds opt-in --cache-memory-fraction 0.5: the server detects its memory allowance after srun launches it and assigns half to cached payloads shared across all datasets and connections.

Linux detection resolves the process's cgroup v1/v2 mount and checks visible ancestor limits, combines those with available Slurm per-node/per-CPU memory information, and bounds the result by physical RAM. A Slurm job with no discoverable limit fails startup instead of assuming it owns the whole node. The detected source and resulting allowance are logged to stderr without disturbing the stdio protocol.

Block and volume-grid caches share admission accounting and reclaim unpinned entries under pressure. Payloads remain charged while requests or escaped shared references retain them, including after dataset closure. Initial block budgets use the detected allowance; subsequent client requests are capped by it. The existing per-dataset volume-grid cap still applies. Omitting the option preserves existing behavior.

Includes a new Slurm workflow guide with a named-allocation launcher, a Riker example, automatic-sizing instructions, cleanup, and troubleshooting, linked from the README and bundled help.

Validation:

• macOS Release builds of the server and Qt client, with compiler warnings treated as errors.
• 17 focused CTest cases passed, including fixtures, memory detection, shared-cache concurrency and allocation-failure rollback, cross-connection reclamation, combined block/grid accounting, the real server over stdio, and existing local query/pipeline regressions.
• Standalone AddressSanitizer/UndefinedBehaviorSanitizer runs passed for the cache and memory-detection tests.
• Bash example syntax, local documentation/resource paths, and whitespace checks passed.

The allowance bounds cached payloads, not process RSS: temporary reads, rendering buffers, particles, and other job processes still require headroom. Detection runs once at startup. Linux hierarchy cases are tested through injected OS inputs; a live Slurm/Riker allocation has not been tested locally.