v1.0.0-rc.1: bench infra + LLM-pluggable corrector dual transport

samkhya-bench: WAVE4-F JOB-Slow real head-to-head harness wired,
cold_cache module via posix_fadvise(POSIX_FADV_DONTNEED) for user-priv
hosts, csv_to_parquet utility, llm_latency + tabpfn_latency campaign
runners, ablation_runner --l4-variant v3 (WAVE5-E recovery), wave5f_audit
example.

samkhya-gpudb/scripts: LLM-pluggable corrector ships dual transport —
canonical Python FastAPI (llm_infer_server.py, port 8766, what file 19
§4.1 measured) and Node TypeScript broader-appeal port
(llm_infer_server.ts, port 8767). Same wire contract (POST /infer + GET
/health, FEATURE_LEN=7, baseline_estimate u64, 4 backends:
anthropic/openai/local/dummy). TabPFN-2.5 inference server
(tabpfn_infer_server.py) for WAVE5-L2 measurement. Run scripts for both
campaigns.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: aiexplore369

samkhya-bench/Cargo.toml

@@ -14,7 +14,7 @@ publish = false
 
 [dependencies]
 samkhya-core = { workspace = true, features = ["gbt", "additive_gbt"] }
-samkhya-datafusion = { path = "../samkhya-datafusion", version = "1.0" }
+samkhya-datafusion = { path = "../samkhya-datafusion", version = "1.0.0-rc.1" }
 clap.workspace = true
 tokio.workspace = true
 datafusion = "46"
@@ -28,3 +28,18 @@ serde_json.workspace = true
 # `GbtCorrector::train` (which hardcodes `baseline_estimate`-only
 # features). Additive — does not affect any other binary.
 gbdt = "0.1"
+# WAVE-5M: cold-cache eviction via posix_fadvise(POSIX_FADV_DONTNEED).
+# libc is already transitively in tree (datafusion → arrow → ...); the
+# explicit dep here makes the cold_cache module's unsafe FFI call
+# auditable. Unix-gated at use site.
+libc = "0.2"
+# WAVE-5L: tabpfn_latency client (file 14). Pure-Rust rustls HTTP client,
+# matches the transport used by samkhya-core::residual::tabpfn so the
+# wire-level overhead is comparable to production. Only the bin needs
+# this; library code is unchanged.
+ureq = { version = "2", default-features = false, features = ["json", "tls"] }
+
+[dev-dependencies]
+# Used by the cold_cache unit test to create a 100 MiB synthetic
+# parquet-shaped file in a self-cleaning temp directory.
+tempfile = "3"

samkhya-bench/examples/wave5f_audit.rs

@@ -0,0 +1,70 @@
+//! WAVE-5F audit helper: read row_count blobs from CSV-built `*.puffin`
+//! and Parquet-built `*.parquet.puffin` sidecars and print a table.
+//!
+//! Throwaway — not part of the v1.0 build matrix. Run via:
+//!   cargo run -p samkhya-bench --release --example wave5f_audit -- samkhya-bench/data/job
+
+use std::fs::File;
+use std::path::Path;
+
+use samkhya_bench::imdb::{ROW_COUNT_KIND, TABLES};
+use samkhya_core::puffin::PuffinReader;
+
+fn read_row_count(path: &Path) -> Option<u64> {
+    let file = File::open(path).ok()?;
+    let mut reader = PuffinReader::open(file).ok()?;
+    let metas = reader.blobs().to_vec();
+    for (i, meta) in metas.iter().enumerate() {
+        if meta.kind == ROW_COUNT_KIND {
+            let payload = reader.read_blob(i).ok()?;
+            if payload.len() >= 8 {
+                let mut buf = [0u8; 8];
+                buf.copy_from_slice(&payload[..8]);
+                return Some(u64::from_le_bytes(buf));
+            }
+        }
+    }
+    None
+}
+
+fn main() {
+    let imdb_dir = std::env::args()
+        .nth(1)
+        .unwrap_or_else(|| "samkhya-bench/data/job".to_string());
+    let imdb_dir = Path::new(&imdb_dir);
+    println!(
+        "{:<20} {:>14} {:>14} {:>10}",
+        "table", "csv_rows", "parquet_rows", "match"
+    );
+    println!("{}", "-".repeat(62));
+    let mut total_csv: u64 = 0;
+    let mut total_parq: u64 = 0;
+    let mut all_match = true;
+    for &t in TABLES {
+        let csv_side = imdb_dir.join(format!("{t}.puffin"));
+        let parq_side = imdb_dir.join(format!("{t}.parquet.puffin"));
+        let c = read_row_count(&csv_side);
+        let p = read_row_count(&parq_side);
+        let m = matches!((c, p), (Some(x), Some(y)) if x == y);
+        if !m {
+            all_match = false;
+        }
+        total_csv += c.unwrap_or(0);
+        total_parq += p.unwrap_or(0);
+        println!(
+            "{:<20} {:>14} {:>14} {:>10}",
+            t,
+            c.map(|n| n.to_string()).unwrap_or_else(|| "MISSING".into()),
+            p.map(|n| n.to_string()).unwrap_or_else(|| "MISSING".into()),
+            if m { "yes" } else { "NO" }
+        );
+    }
+    println!("{}", "-".repeat(62));
+    println!(
+        "{:<20} {:>14} {:>14} {:>10}",
+        "TOTAL",
+        total_csv,
+        total_parq,
+        if all_match { "yes" } else { "NO" }
+    );
+}