fix(apr-cli): build without the inference feature (#4972)

crates/apr-cli/src/commands/chat.rs

@@ -40,6 +40,7 @@ use aprender::serialization::apr::AprReader;
 // asking…` on Qwen3-1.7B. Two functions with the same name, two enums, one of
 // which could not express the right answer.
 use colored::Colorize;
+#[cfg(feature = "inference")]
 use realizar::chat_template::{
     auto_detect_template, detect_format_from_name, ChatMessage, ChatTemplateEngine, TemplateFormat,
 };
@@ -682,7 +683,9 @@ fn print_welcome_banner(path: &Path, config: &ChatConfig) {
     print_welcome_banner_for(path, detect_format(path), config);
 }
 
-fn print_welcome_banner_for(path: &Path, format: ModelFormat, config: &ChatConfig) {
+/// The chat-template family the banner names, detected from the file name.
+#[cfg(feature = "inference")]
+fn banner_template_family(path: &Path) -> &'static str {
     // Detect chat template format from model name (Toyota Way: Visual Control)
     let model_name = path
         .file_stem()
@@ -692,7 +695,18 @@ fn print_welcome_banner_for(path: &Path, format: ModelFormat, config: &ChatConfi
     // #3801: ONE spelling of the name. This was a second copy of
     // `template_format_name`, and two copies of a match over an enum are how a
     // new variant gets handled in one place and not the other.
-    let template_name = crate::chat::realizar_chat::template_format_name(template_format);
+    crate::chat::realizar_chat::template_format_name(template_format)
+}
+
+/// #4972: without `inference` there is no template detector, and the session
+/// refuses before its first turn.
+#[cfg(not(feature = "inference"))]
+fn banner_template_family(_path: &Path) -> &'static str {
+    "none: this apr was built without the `inference` feature"
+}
+
+fn print_welcome_banner_for(path: &Path, format: ModelFormat, config: &ChatConfig) {
+    let template_name = banner_template_family(path);
 
     match format {
         ModelFormat::Apr => {

crates/apr-cli/src/commands/chat_generate_session.rs

@@ -10,7 +10,7 @@ struct ChatSession;
 
 #[cfg(not(feature = "inference"))]
 impl ChatSession {
-    fn new(_path: &Path) -> Result<Self, CliError> {
+    fn new(_path: &Path, _force_cpu: bool) -> Result<Self, CliError> {
         Err(CliError::ValidationFailed(
             "Chat requires the 'inference' feature (realizar). Rebuild with: \
              cargo install --path crates/apr-cli --features inference"

crates/apr-cli/src/commands/golden_output.rs

@@ -355,6 +355,7 @@ fn thinking_on_budget_override() -> std::result::Result<Option<(usize, String)>,
 /// One matched row of the `models` table. A row that declares no `budget`
 /// REFUSES rather than inheriting the default, which was measured on a
 /// different model (#3907).
+#[cfg(feature = "inference")]
 fn thinking_on_budget_row(
     v: &serde_yaml::Value,
     pat: &str,
@@ -385,6 +386,7 @@ fn thinking_on_budget_row(
 
 /// The first `models` pattern that matches `model_file`, or `None` when the
 /// model is unlisted and the `default` row applies.
+#[cfg(feature = "inference")]
 fn thinking_on_budget_match(
     doc: &serde_yaml::Value,
     model_file: &str,
@@ -400,6 +402,7 @@ fn thinking_on_budget_match(
 }
 
 /// The `default` row, for a model no pattern names.
+#[cfg(feature = "inference")]
 fn thinking_on_budget_default(
     doc: &serde_yaml::Value,
 ) -> std::result::Result<(usize, String), String> {
@@ -418,6 +421,7 @@ fn thinking_on_budget_default(
     Ok((usize::try_from(b).unwrap_or(0), format!("default: {basis}")))
 }
 
+#[cfg(feature = "inference")]
 fn thinking_on_budget_for(model_file: &str) -> std::result::Result<(usize, String), String> {
     if let Some(over) = thinking_on_budget_override()? {
         return Ok(over);

crates/apr-cli/src/commands/output_verification.rs

@@ -1055,6 +1055,7 @@ fn run_golden_output_gate_runtime(
 /// pass message's ON-leg clause (empty when the model has no thinking-on case), and
 /// `Err(failed)` ends the gate. Extracted from `run_golden_output_gate_runtime` unchanged,
 /// to keep that function under the complexity ratchet.
+#[cfg(feature = "inference")]
 fn runtime_thinking_on_leg(
     path: &Path,
     architecture: Option<&str>,

crates/apr-cli/src/commands/run_entry.rs

@@ -212,6 +212,7 @@ fn reconcile_and_emit(
 /// (not a GGUF, a hub id rather than a path, an unreadable header), and an
 /// unknown architecture is never refused — this predicate lists what we know we
 /// did NOT build, never what we support.
+#[cfg(feature = "inference")]
 pub(crate) fn forced_accelerator_refusal(
     accel_forced: bool,
     architecture: Option<&str>,
@@ -222,6 +223,20 @@ pub(crate) fn forced_accelerator_refusal(
     realizar::capability::no_cuda_forward_reason(architecture?)
 }
 
+/// #4972: a build without `inference` has no forward on any accelerator, so a
+/// forced one is refused whatever the architecture, including an unread one.
+#[cfg(not(feature = "inference"))]
+pub(crate) fn forced_accelerator_refusal(
+    accel_forced: bool,
+    _architecture: Option<&str>,
+) -> Option<String> {
+    accel_forced.then(|| {
+        "an accelerator was forced, but this apr was built without the `inference` \
+         feature, so it has no forward on any device (#4972)"
+            .to_string()
+    })
+}
+
 /// Read the declared architecture from a model path without loading it.
 ///
 /// **A bounded PREFIX read, deliberately not an mmap.** The first draft called

crates/apr-cli/src/commands/run_tests_accel_reconcile.rs

@@ -197,7 +197,7 @@ fn json_plus_benchmark_is_not_a_machine_surface() {
 // #3817: --gpu on an architecture with no CUDA forward refuses BEFORE the load
 // =============================================================================
 
-#[cfg(test)]
+#[cfg(all(test, feature = "inference"))]
 mod forced_accelerator_refusal_tests {
     use crate::commands::run::forced_accelerator_refusal;
 
@@ -280,6 +280,29 @@ mod forced_accelerator_refusal_tests {
     }
 }
 
+/// #4972: without `inference` there is no forward on any device, so a forced
+/// accelerator is refused for every architecture, an unread one included, and
+/// an unforced run is never refused.
+#[cfg(all(test, not(feature = "inference")))]
+mod forced_accelerator_refusal_without_inference_tests {
+    use crate::commands::run::forced_accelerator_refusal;
+
+    #[test]
+    fn a_forced_accelerator_is_refused_for_every_architecture() {
+        for arch in [Some("qwen2"), Some("llama"), Some("qwen35moe"), None] {
+            let reason = forced_accelerator_refusal(true, arch)
+                .expect("a build without inference has no forward to reach");
+            assert!(reason.contains("inference"), "{reason}");
+        }
+    }
+
+    #[test]
+    fn an_unforced_run_is_not_refused() {
+        assert!(forced_accelerator_refusal(false, Some("qwen2")).is_none());
+        assert!(forced_accelerator_refusal(false, None).is_none());
+    }
+}
+
 // ── #3826: `fell_back` must name a fallback that HAPPENED ──────────────────
 //
 // `apr run --format json` reported `"fell_back": false` on a run whose own

crates/apr-cli/src/commands/serve/mod.rs

@@ -9,6 +9,7 @@ pub mod auth;
 pub mod handlers;
 #[cfg(feature = "inference")]
 pub mod ollama;
+#[cfg(feature = "inference")]
 pub(crate) mod route_index;
 pub mod routes;
 #[cfg(feature = "inference")]

crates/apr-cli/src/dispatch_analysis.rs

@@ -495,6 +495,7 @@ fn dispatch_analysis_commands_rest(cli: &Cli) -> Option<Result<(), CliError>> {
         // complexity debt the pre-commit gate refuses to let any edit ride on.
         #[cfg(feature = "inference")]
         ExtendedCommands::Capability { json } => commands::capability::run(*json || cli.json),
+        #[cfg(feature = "inference")]
         ExtendedCommands::Devices { json } => commands::devices::run(*json || cli.json),
         ExtendedCommands::OtlpLint {
             otlp_file,

docs/audits/surface_audit.csv

@@ -117,7 +117,7 @@ apr,apr cgp profile wgpu,6,x86_64-linux,NONE,no,9,profiling,crates/aprender-cgp/
 apr,apr cgp roofline,6,x86_64-linux,NONE,no,7,qa-cgp,crates/aprender-cgp/src/cli.rs:64,low
 apr,apr cgp trace,6,x86_64-linux,NONE,no,7,qa-cgp,crates/aprender-cgp/src/cli.rs:99,low
 apr,apr cgp tui,6,x86_64-linux,NONE,no,7,qa-cgp,crates/aprender-cgp/src/cli.rs:115,low
-apr,apr chat,6,x86_64-linux,llama.cpp,yes,1,apr-lint-diag,crates/apr-cli/src/extended_commands.rs:40 (ACCELERATOR_GUARD crates/apr-cli/src/dispatch_analysis.rs:1710),high
+apr,apr chat,6,x86_64-linux,llama.cpp,yes,1,apr-lint-diag,crates/apr-cli/src/extended_commands.rs:40 (ACCELERATOR_GUARD crates/apr-cli/src/dispatch_analysis.rs:1722),high
 apr,apr chat --backend cpu,6,x86_64-linux,ollama,no,1,apr-lint-diag,crates/apr-cli/src/extended_commands.rs:85 (BACKEND_VALUES crates/apr-cli/src/commands_enum.rs:16),high
 apr,apr chat --backend cuda,6,x86_64-linux;nvidia-cuda,ollama,no,1,apr-lint-diag,crates/apr-cli/src/extended_commands.rs:85 (BACKEND_VALUES crates/apr-cli/src/commands_enum.rs:16),medium
 apr,apr chat --backend wgpu,6,UNKNOWN,ollama,no,1,apr-lint-diag,crates/apr-cli/src/extended_commands.rs:85 (BACKEND_VALUES crates/apr-cli/src/commands_enum.rs:16),high
@@ -184,16 +184,16 @@ apr,apr embed-viz-lint,1,x86_64-linux,NONE,yes,1,apr-lint-diag,crates/apr-cli/sr
 apr,apr embeddings-lint,1,x86_64-linux,NONE,yes,1,apr-lint-diag,crates/apr-cli/src/extended_commands.rs:1173,high
 apr,apr encrypt,4,x86_64-linux,NONE,yes,10,apr-core-commands,crates/apr-cli/src/tool_commands.rs:121,high
 apr,apr eval,6,x86_64-linux,llama.cpp,yes,1,apr-lint-diag,crates/apr-cli/src/extended_commands.rs:124,high
-apr,apr eval --task classify,6,x86_64-linux,scikit-learn,no,8,rag-eval,crates/apr-cli/src/dispatch_analysis.rs:1406 (flag crates/apr-cli/src/extended_commands.rs:145),high
-apr,apr eval --task code,6,x86_64-linux,NONE,no,8,rag-eval,crates/apr-cli/src/dispatch_analysis.rs:1420 (flag crates/apr-cli/src/extended_commands.rs:145),high
-apr,apr eval --task compare,6,x86_64-linux,NONE,no,8,rag-eval,crates/apr-cli/src/dispatch_analysis.rs:1444 (flag crates/apr-cli/src/extended_commands.rs:145),high
-apr,apr eval --task contamination,6,x86_64-linux,huggingface-transformers,no,8,rag-eval,crates/apr-cli/src/dispatch_analysis.rs:1441 (flag crates/apr-cli/src/extended_commands.rs:145),high
-apr,apr eval --task correlation,6,x86_64-linux,NONE,no,8,rag-eval,crates/apr-cli/src/dispatch_analysis.rs:1446 (flag crates/apr-cli/src/extended_commands.rs:145),high
-apr,apr eval --task human,6,x86_64-linux,NONE,no,8,rag-eval,crates/apr-cli/src/dispatch_analysis.rs:1447 (flag crates/apr-cli/src/extended_commands.rs:145),high
-apr,apr eval --task humaneval,6,x86_64-linux,NONE,no,8,rag-eval,crates/apr-cli/src/dispatch_analysis.rs:1423 (flag crates/apr-cli/src/extended_commands.rs:145),high
-apr,apr eval --task mbpp,6,x86_64-linux,NONE,no,8,rag-eval,crates/apr-cli/src/dispatch_analysis.rs:1432 (flag crates/apr-cli/src/extended_commands.rs:145),high
-apr,apr eval --task plan,6,x86_64-linux,NONE,no,8,rag-eval,crates/apr-cli/src/dispatch_analysis.rs:1448 (flag crates/apr-cli/src/extended_commands.rs:145),high
-apr,apr eval --task verify,6,x86_64-linux,NONE,no,8,rag-eval,crates/apr-cli/src/dispatch_analysis.rs:1445 (flag crates/apr-cli/src/extended_commands.rs:145),high
+apr,apr eval --task classify,6,x86_64-linux,scikit-learn,no,8,rag-eval,crates/apr-cli/src/dispatch_analysis.rs:1441 (flag crates/apr-cli/src/extended_commands.rs:145),high
+apr,apr eval --task code,6,x86_64-linux,NONE,no,8,rag-eval,crates/apr-cli/src/dispatch_analysis.rs:1455 (flag crates/apr-cli/src/extended_commands.rs:145),high
+apr,apr eval --task compare,6,x86_64-linux,NONE,no,8,rag-eval,crates/apr-cli/src/dispatch_analysis.rs:1479 (flag crates/apr-cli/src/extended_commands.rs:145),high
+apr,apr eval --task contamination,6,x86_64-linux,huggingface-transformers,no,8,rag-eval,crates/apr-cli/src/dispatch_analysis.rs:1476 (flag crates/apr-cli/src/extended_commands.rs:145),high
+apr,apr eval --task correlation,6,x86_64-linux,NONE,no,8,rag-eval,crates/apr-cli/src/dispatch_analysis.rs:1481 (flag crates/apr-cli/src/extended_commands.rs:145),high
+apr,apr eval --task human,6,x86_64-linux,NONE,no,8,rag-eval,crates/apr-cli/src/dispatch_analysis.rs:1482 (flag crates/apr-cli/src/extended_commands.rs:145),high
+apr,apr eval --task humaneval,6,x86_64-linux,NONE,no,8,rag-eval,crates/apr-cli/src/dispatch_analysis.rs:1458 (flag crates/apr-cli/src/extended_commands.rs:145),high
+apr,apr eval --task mbpp,6,x86_64-linux,NONE,no,8,rag-eval,crates/apr-cli/src/dispatch_analysis.rs:1467 (flag crates/apr-cli/src/extended_commands.rs:145),high
+apr,apr eval --task plan,6,x86_64-linux,NONE,no,8,rag-eval,crates/apr-cli/src/dispatch_analysis.rs:1483 (flag crates/apr-cli/src/extended_commands.rs:145),high
+apr,apr eval --task verify,6,x86_64-linux,NONE,no,8,rag-eval,crates/apr-cli/src/dispatch_analysis.rs:1480 (flag crates/apr-cli/src/extended_commands.rs:145),high
 apr,apr experiment view,6,x86_64-linux,pytorch,no,1,apr-lint-diag,crates/apr-cli/src/extended_commands.rs:1568,low
 apr,apr explain,6,x86_64-linux,safetensors,yes,10,apr-core-commands,crates/apr-cli/src/commands_enum.rs:447,high
 apr,apr explain-token-lint,1,x86_64-linux,NONE,yes,1,apr-lint-diag,crates/apr-cli/src/extended_commands.rs:1066,low
@@ -239,9 +239,9 @@ apr,apr oom-lint,1,x86_64-linux,NONE,yes,1,apr-lint-diag,crates/apr-cli/src/exte
 apr,apr oracle,9,x86_64-linux,NONE,yes,10,apr-core-commands,crates/apr-cli/src/tool_commands.rs:84,medium
 apr,apr otlp-lint,1,x86_64-linux,NONE,yes,1,apr-lint-diag,crates/apr-cli/src/extended_commands.rs:1108,high
 apr,apr parity,9,x86_64-linux,NONE,yes,1,apr-lint-diag,crates/apr-cli/src/extended_commands.rs:317,medium
-apr,apr parity --per-op,9,"x86_64-linux,aarch64-linux",NONE,no,1,apr-lint-diag,crates/apr-cli/src/extended_commands.rs:330 (dispatch crates/apr-cli/src/dispatch_analysis.rs:1516),high
-apr,apr parity --out,7,"x86_64-linux,aarch64-linux",NONE,no,1,apr-lint-diag,crates/apr-cli/src/extended_commands.rs:333 (dispatch crates/apr-cli/src/dispatch_analysis.rs:1517),high
-apr,apr parity --threshold,7,"x86_64-linux,aarch64-linux",NONE,no,1,apr-lint-diag,crates/apr-cli/src/extended_commands.rs:336 (dispatch crates/apr-cli/src/dispatch_analysis.rs:1517),high
+apr,apr parity --per-op,9,"x86_64-linux,aarch64-linux",NONE,no,1,apr-lint-diag,crates/apr-cli/src/extended_commands.rs:330 (dispatch crates/apr-cli/src/dispatch_analysis.rs:1564),high
+apr,apr parity --out,7,"x86_64-linux,aarch64-linux",NONE,no,1,apr-lint-diag,crates/apr-cli/src/extended_commands.rs:333 (dispatch crates/apr-cli/src/dispatch_analysis.rs:1564),high
+apr,apr parity --threshold,7,"x86_64-linux,aarch64-linux",NONE,no,1,apr-lint-diag,crates/apr-cli/src/extended_commands.rs:336 (dispatch crates/apr-cli/src/dispatch_analysis.rs:1564),high
 apr,apr pipeline apply,6,x86_64-linux,huggingface-transformers,no,10,apr-core-commands,crates/apr-cli/src/pipeline_commands.rs:28,low
 apr,apr pipeline plan,6,x86_64-linux,huggingface-transformers,no,10,apr-core-commands,crates/apr-cli/src/pipeline_commands.rs:9,low
 apr,apr pipeline status,6,x86_64-linux,huggingface-transformers,no,10,apr-core-commands,crates/apr-cli/src/pipeline_commands.rs:51,low