perf: route whisper + llm n_threads through physical-core helper

Both inference call sites previously called `num_cpus::get()` (logical
thread count). Established whisper.cpp / llama.cpp guidance is that
SMT siblings contend for shared FPU resources during heavy F16/F32
matmul, so going past physical core count anti-scales. Empirical
sweep on Whisper Tiny / 11s JFK clip / Ryzen 5 4650U (6c12t):

  n_threads | xc_time | RTF    | speedup_vs_1
  ----------|---------|--------|-------------
          1 |   0.33s |  0.030 |   1.00x
          2 |   0.33s |  0.030 |   1.00x
          4 |   0.30s |  0.028 |   1.09x
          6 |   0.32s |  0.029 |   1.04x
          8 |   0.31s |  0.028 |   1.06x
         12 |   0.32s |  0.029 |   1.03x

Tiny doesn't scale (work dominated by overhead) but the larger
Whisper variants and Qwen LLMs do. Sources: whisper.cpp #200, #1033,
#1252, #403; llama.cpp #3167, #572.

Changes:
- crates/core/src/constants.rs:
  * MIN_INFERENCE_THREADS lowered 4 → 2 (research-derived floor)
  * MAX_INFERENCE_THREADS = 8 added (research-derived ceiling)
  * inference_thread_count() rewritten:
      - reads MAGNOTIA_INFERENCE_THREADS env var (override)
      - num_cpus::get_physical() with available_parallelism fallback
      - clamped to [MIN_INFERENCE_THREADS, MAX_INFERENCE_THREADS]
- crates/core/Cargo.toml: + num_cpus = "1"
- crates/transcription/src/whisper_rs_backend.rs: call site uses helper.
- crates/llm/src/lib.rs: call site uses helper.
- crates/transcription/Cargo.toml: drop num_cpus from [features] +
  [dependencies] (production no longer needs it). Move to
  [dev-dependencies] for tests/thread_sweep.rs only.
- crates/llm/Cargo.toml: drop num_cpus = "1" (no longer used directly).

Per-machine maps (after this patch):
  Ryzen 5 4650U  6c12t → 6
  big-iron 12c24t       → 8 (clamp; users can override)
  cheap 2c2t laptop     → 2
  1c container/VM       → 2

Adds crates/transcription/tests/thread_sweep.rs — env-gated like
jfk_bench, prints the table above against any model + WAV. Useful for
re-baselining on new hardware or when tuning the clamp values.

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

crates/core/Cargo.toml

@@ -10,3 +10,4 @@ serde_json = "1"
 thiserror = "2"
 sysinfo = "0.35"
 async-trait = "0.1"
+num_cpus = "1"

crates/core/src/constants.rs

@@ -18,8 +18,21 @@ pub const MIN_CHUNK_SAMPLES: usize = 8000;
 /// Post-processing thresholds.
 pub const SMART_PARAGRAPH_GAP_SECS: f64 = 2.0;
 
-/// Thread count for inference. Leaves headroom for the UI thread.
-pub const MIN_INFERENCE_THREADS: usize = 4;
+/// Lower bound for inference threads. Single-threaded inference is
+/// measurably worse than two on every multi-core part; we never go below
+/// 2. (Whisper Tiny is the exception where the work is so small that
+/// thread count barely matters — but the floor still costs nothing.)
+pub const MIN_INFERENCE_THREADS: usize = 2;
+
+/// Upper bound for inference threads. Both whisper.cpp and llama.cpp
+/// scaling flattens around physical core count; SMT siblings contend
+/// for shared FPU resources during heavy F16/F32 matmul, which means
+/// going past physical cores often anti-scales. Empirical evidence:
+/// whisper.cpp issue #200 (sweet spot at 7t on 8c/16t Ryzen 3700X),
+/// llama.cpp #3167 ("SMT hurts inference"). 8 is a conservative
+/// ceiling that leaves <5% on the table for big-iron desktops while
+/// keeping consumer 6c/12t laptops out of contention territory.
+pub const MAX_INFERENCE_THREADS: usize = 8;
 
 /// History limits.
 pub const HISTORY_MAX_ENTRIES: usize = 100;
@@ -40,10 +53,37 @@ pub const VAD_SPEECH_PAD_MS: u32 = 100;
 /// Model download chunk size for progress reporting.
 pub const DOWNLOAD_CHUNK_BYTES: usize = 65_536;
 
-/// Inference thread count based on available parallelism.
+/// Inference thread count, clamped to physical-core budget.
+///
+/// Returns the system's physical-core count, clamped to
+/// `[MIN_INFERENCE_THREADS, MAX_INFERENCE_THREADS]`. Uses physical
+/// rather than logical cores because SMT/hyperthreaded siblings
+/// contend for shared FPU resources during heavy matmul; counting
+/// them as additional workers is well-documented to anti-scale on
+/// both whisper.cpp and llama.cpp.
+///
+/// Falls back to `available_parallelism` only if the physical-core
+/// probe is unavailable (some non-Linux/non-Windows platforms or
+/// containerised environments).
+///
+/// Users can override at runtime by setting
+/// `MAGNOTIA_INFERENCE_THREADS=N` — useful for benchmarking and for
+/// users on big-iron desktops who want to push past the cap.
 pub fn inference_thread_count() -> usize {
-    std::thread::available_parallelism()
-        .map(|p| p.get().saturating_sub(1))
-        .unwrap_or(MIN_INFERENCE_THREADS)
-        .max(MIN_INFERENCE_THREADS)
+    if let Ok(s) = std::env::var("MAGNOTIA_INFERENCE_THREADS") {
+        if let Ok(n) = s.parse::<usize>() {
+            if n > 0 {
+                return n;
+            }
+        }
+    }
+    let physical = num_cpus::get_physical();
+    let chosen = if physical > 0 {
+        physical
+    } else {
+        std::thread::available_parallelism()
+            .map(|p| p.get())
+            .unwrap_or(MIN_INFERENCE_THREADS)
+    };
+    chosen.clamp(MIN_INFERENCE_THREADS, MAX_INFERENCE_THREADS)
 }

crates/llm/Cargo.toml

@@ -20,7 +20,6 @@ magnotia-core = { path = "../core" }
 encoding_rs = "0.8"
 futures-util = "0.3"
 llama-cpp-2 = { version = "0.1.144", default-features = false }
-num_cpus = "1"
 reqwest = { version = "0.12", default-features = false, features = ["rustls-tls", "stream"] }
 serde = { version = "1", features = ["derive"] }
 serde_json = "1"

crates/llm/src/lib.rs

@@ -163,7 +163,8 @@ impl LlmEngine {
         }
 
         let n_ctx = preflight_context_window(prompt_tokens.len(), config.max_tokens)?;
-        let thread_count = i32::try_from(num_cpus::get().max(1)).unwrap_or(4);
+        let thread_count = i32::try_from(magnotia_core::constants::inference_thread_count())
+            .unwrap_or(4);
         let ctx_params = LlamaContextParams::default()
             .with_n_ctx(Some(
                 NonZeroU32::new(n_ctx).expect("n_ctx must be non-zero"),

crates/transcription/Cargo.toml

@@ -17,7 +17,7 @@ build = "build.rs"
 # but skip the Vulkan backend. Build CPU-only with:
 #   cargo build -p magnotia-transcription --no-default-features --features whisper
 default = ["whisper", "whisper-vulkan"]
-whisper = ["dep:whisper-rs", "dep:num_cpus"]
+whisper = ["dep:whisper-rs"]
 whisper-vulkan = ["whisper-rs?/vulkan"]
 
 [dependencies]
@@ -40,10 +40,6 @@ sha2 = "0.10"
 # additive via the `whisper-vulkan` feature so non-GPU targets can drop it.
 whisper-rs = { version = "0.16", default-features = false, optional = true }
 
-# Direct whisper-rs backend (WhisperRsBackend): thread pool sizing.
-# Gated alongside whisper-rs since no other code in this crate needs it.
-num_cpus = { version = "1", optional = true }
-
 # Typed error enum used by WhisperRsBackend + elsewhere. Kept
 # unconditional because it is a derive-macro crate with negligible
 # build cost.
@@ -56,3 +52,7 @@ tracing = "0.1"
 # TcpListener fixture for the download resume tests (mirrors magnotia-llm).
 tokio = { version = "1", features = ["rt", "sync", "net", "io-util", "macros"] }
 tempfile = "3"
+# Test-only — used by tests/thread_sweep.rs to label physical vs logical
+# core counts in the scaling table. Production code uses the
+# `magnotia_core::constants::inference_thread_count` helper instead.
+num_cpus = "1"

crates/transcription/src/whisper_rs_backend.rs

@@ -10,6 +10,7 @@ use std::path::Path;
 
 use whisper_rs::{FullParams, SamplingStrategy, WhisperContext, WhisperContextParameters};
 
+use magnotia_core::constants::inference_thread_count;
 use magnotia_core::error::{MagnotiaError, Result};
 use magnotia_core::types::{Segment, TranscriptionOptions};
 
@@ -77,7 +78,7 @@ impl Transcriber for WhisperRsBackend {
                 params.set_initial_prompt(prompt);
             }
         }
-        params.set_n_threads(num_cpus::get() as i32);
+        params.set_n_threads(inference_thread_count() as i32);
         params.set_print_special(false);
         params.set_print_progress(false);
         params.set_print_realtime(false);

crates/transcription/tests/thread_sweep.rs

@@ -0,0 +1,76 @@
+//! Thread-count scaling sweep for Whisper Tiny.
+//! Runs the JFK clip at n_threads = 1, 2, 4, 6, 8, 12, prints RTF table.
+//! Gated on the same env vars as jfk_bench.
+
+use std::env;
+use std::time::Instant;
+
+#[test]
+fn whisper_thread_count_sweep() {
+    let Ok(model_path) = env::var("MAGNOTIA_WHISPER_TEST_MODEL") else { return };
+    let Ok(audio_path) = env::var("MAGNOTIA_WHISPER_TEST_AUDIO") else { return };
+
+    use whisper_rs::{FullParams, SamplingStrategy, WhisperContext, WhisperContextParameters};
+
+    let bytes = std::fs::read(&audio_path).expect("read wav");
+    let sample_rate = u32::from_le_bytes(bytes[24..28].try_into().unwrap());
+    let pcm = &bytes[44..];
+    let samples: Vec<f32> = pcm
+        .chunks_exact(2)
+        .map(|c| i16::from_le_bytes([c[0], c[1]]) as f32 / 32768.0)
+        .collect();
+    let audio_secs = samples.len() as f64 / sample_rate as f64;
+    eprintln!("[sweep] audio: {:.2}s @ {} Hz", audio_secs, sample_rate);
+
+    let logical = num_cpus::get();
+    let physical = num_cpus::get_physical();
+    eprintln!("[sweep] CPU: physical={}, logical={}", physical, logical);
+
+    let ctx = WhisperContext::new_with_params(&model_path, WhisperContextParameters::default())
+        .expect("model load");
+
+    // Warm-up pass at default to prime caches
+    {
+        let mut state = ctx.create_state().expect("state");
+        let mut params = FullParams::new(SamplingStrategy::Greedy { best_of: 1 });
+        params.set_language(Some("en"));
+        params.set_n_threads(physical as i32);
+        params.set_print_special(false);
+        params.set_print_progress(false);
+        params.set_print_realtime(false);
+        state.full(params, &samples).expect("warmup");
+    }
+
+    let mut targets: Vec<i32> = vec![1, 2, 4, physical as i32, logical as i32];
+    if logical >= 8 && !targets.contains(&8) { targets.push(8); }
+    targets.sort();
+    targets.dedup();
+
+    eprintln!("");
+    eprintln!("=== n_threads scaling ===");
+    eprintln!("n_threads | xc_time | RTF    | speedup_vs_1");
+    eprintln!("----------|---------|--------|-------------");
+    let mut baseline_dur: Option<f64> = None;
+    for n in &targets {
+        // Two runs, take the min (deeper of L2/L3 effects; we want best-case)
+        let mut best = f64::MAX;
+        for _ in 0..2 {
+            let mut state = ctx.create_state().expect("state");
+            let mut params = FullParams::new(SamplingStrategy::Greedy { best_of: 1 });
+            params.set_language(Some("en"));
+            params.set_n_threads(*n);
+            params.set_print_special(false);
+            params.set_print_progress(false);
+            params.set_print_realtime(false);
+            let t = Instant::now();
+            state.full(params, &samples).expect("transcribe");
+            let dur = t.elapsed().as_secs_f64();
+            if dur < best { best = dur; }
+        }
+        let rtf = best / audio_secs;
+        let speedup = baseline_dur.map(|b| b / best).unwrap_or(1.0);
+        if baseline_dur.is_none() { baseline_dur = Some(best); }
+        eprintln!("{:>9} | {:>6.2}s | {:>6.3} | {:>6.2}x",
+            n, best, rtf, speedup);
+    }
+}