perf: remove memory bottleneck in stages 5 and 6

CLAUDE.md

@@ -106,8 +106,8 @@ ProofTranscript: Transcript               — Fiat-Shamir transcript (Blake2bTra
 
 - `DensePolynomial<F>`: Full field-element coefficients
 - `CompactPolynomial<T>`: Small scalar coefficients (u8–i128), promoted to field on bind
-- `RaPolynomial`: Lazy materialization via Round1→Round2→Round3→RoundN state machine
-- `SharedRaPolynomials`: Shares eq tables across N polynomials for memory efficiency
+- `RaPolynomial`: Lazy materialization via table-doubling state machine (TableRound → RoundN, materializes at 8 groups)
+- `SharedRaPolynomials`: Shares eq tables across N polynomials via table-doubling (TableRound → RoundN, materializes at 16 groups)
 - `PrefixSuffixDecomposition`: Splits polynomial as `Σ P_i(prefix) · Q_i(suffix)` for efficient sumcheck
 - `MultilinearPolynomial<F>`: Enum dispatching over all scalar types + OneHot/RLC variants
 

Cargo.toml

@@ -211,6 +211,7 @@ num-traits = { version = "0.2.19", default-features = false }
 sysinfo = "0.38"
 memory-stats = { version = "1.0.0", features = ["always_use_statm"] }
 mimalloc = "0.1"
+tikv-jemallocator = { version = "0.6", features = ["unprefixed_malloc_on_supported_platforms"] }
 
 # Parallel Processing
 rayon = { version = "^1.8.0" }

jolt-core/Cargo.toml

@@ -38,6 +38,7 @@ prover = [
     "dory/backends",
     "dory/cache", # This includes `parallel` feature
     "dory/disk-persistence",
+    "dep:tikv-jemallocator",
 ]
 # This is for building jolt-core without prover capabilities, e.g. for recursion
 # jolt-core needs std and rayon to compile, so these are the minimal set of features.
@@ -89,6 +90,7 @@ jolt-inlines-keccak256 = { workspace = true, features = [
     "host",
 ], optional = true }
 sysinfo = { workspace = true, optional = true }
+tikv-jemallocator = { workspace = true, optional = true }
 pprof = { version = "0.15", features = [
     "prost-codec",
     "flamegraph",

jolt-core/src/bin/jolt_core.rs

@@ -1,3 +1,10 @@
+#[global_allocator]
+static GLOBAL: tikv_jemallocator::Jemalloc = tikv_jemallocator::Jemalloc;
+
+#[allow(non_upper_case_globals)]
+#[unsafe(export_name = "malloc_conf")]
+pub static malloc_conf: &[u8] = b"dirty_decay_ms:0,muzzy_decay_ms:0\0";
+
 use clap::{Args, Parser, Subcommand, ValueEnum};
 
 #[path = "../../benches/e2e_profiling.rs"]