Add backend performance workflow and tuning

.github/workflows/main.yml

@@ -60,28 +60,33 @@ jobs:
       run: |
         cp Cargo.lock.msrv Cargo.lock
     - run: |
-        cargo test --no-default-features --features rust-backend
+        cargo test
         cargo clean -p liblzma -p liblzma-sys
     - run: |
         cargo test --features parallel
         cargo clean -p liblzma -p liblzma-sys
     - run: |
-        cargo test --features bindgen
+        cargo test --no-default-features --features xz-sys
+        cargo clean -p liblzma -p xz-sys
+    - run: |
+        cargo test --no-default-features --features liblzma-sys
         cargo clean -p liblzma -p liblzma-sys
     # thin-lto and fat-lto feature required use clang as a linker, but linux default linker is gcc. so exclude thin-lto and fat-lto.
     - run: |
-        cargo test --features parallel,static
+        cargo test --no-default-features --features liblzma-sys,parallel,static
         cargo clean -p liblzma -p liblzma-sys
       if: ${{ startsWith(matrix.os, 'ubuntu') }}
     - run: |
-        cargo test --all-features
+        cargo test --test sys_equivalence
         cargo clean -p liblzma -p liblzma-sys
       if: ${{ !startsWith(matrix.os, 'ubuntu') }}
     - run: |
-        cargo run --manifest-path systest/Cargo.toml
+        cargo run --manifest-path systest/Cargo.toml --no-default-features --features xz-sys
+        cargo clean -p xz-sys
+        cargo run --manifest-path systest/Cargo.toml --no-default-features --features liblzma-sys
         cargo clean -p liblzma -p liblzma-sys
       if: matrix.static == 'yes'
-    - run: cargo run --manifest-path systest/Cargo.toml --features bindgen
+    - run: cargo run --manifest-path systest/Cargo.toml --no-default-features --features liblzma-sys,bindgen
       if: matrix.static == 'yes'
 
 

AGENTS.md

@@ -0,0 +1,20 @@
+# AGENTS
+
+## Performance Work
+
+Use the dedicated backend-performance workflow in [`docs/performance-workflow.md`](docs/performance-workflow.md).
+
+Key entry points:
+
+- Full deterministic regression gate: `scripts/compare_backends.sh`
+- Focused backend workloads: `scripts/compare_workloads.sh`
+- High-level API workloads: `scripts/compare_api_workloads.sh`
+- Single-backend profiling: `scripts/profile_backend.sh`
+- Optimized code inspection: `scripts/inspect_codegen.sh`
+
+Important constraints:
+
+- Never compare C and Rust sys backends in the same process. They export the same `lzma_*` symbols.
+- Treat `scripts/compare_backends.sh` as the coarse deterministic gate.
+- Use `qc` and `size` focused workloads for paths that are property-test-like or otherwise input-sensitive.
+- Keep workload shape, iteration count, and profiler command stable while iterating on a hotspot.