M4: Embedded Performance and Scale for v0.5.x

Increase GraphForge practical scale and throughput while preserving the embedded-first, Rust-owned execution model.

In scope:

• deterministic performance baselines and thread-scaling evidence;
• streaming and partitioned Parquet execution with bounded memory and spill;
• CSR-native graph representations and bounded Arrow result shaping;
• bounded process-level compute policy and deterministic CPU-parallel kernels;
• an optional in-process GPU acceleration spike gated by end-to-end parity and measured benefit.

Out of scope:

• a required server, distributed database, or multi-node execution authority;
• Python, Node, NetworkX, igraph, or other foreign graph engines as runtime fallbacks;
• weakening deterministic ordering, cancellation, resource limits, errors, or correctness contracts;
• promising a graph-size ceiling before reproducible evidence exists.

Close only when the canonical M4 closure tracker and all milestone issues are closed, entry and exit performance evidence is reproducible, exact-head CI is green for every changed surface, CPU-only embedded operation remains the universal default, and any shipped accelerator meets its explicit parity and end-to-end performance gate.

Recovery note (2026-08-14):

• M4 was reopened after audit found that the accepted exit ledger predates the later #498/#499–#588 expansion and that #336/#338 still lack their required large-scale outcome evidence.
• #345 must reconcile the actual post-#498 final tree and repair the stale/broken exit ledger before #335 or this milestone can close again.
• M5 (#735; milestone 5) is the separate forward-looking >=1-billion-live-edge and hardened-interchange program. M5 does not waive, replace, or retroactively satisfy M4 acceptance criteria.