gstoner · gstoner · Jul 10, 2026 · Jul 10, 2026 · Jul 10, 2026 · Jul 10, 2026
diff --git a/docs/audit/generated/apple_target_map.csv b/docs/audit/generated/apple_target_map.csv
@@ -98,6 +98,7 @@ load_balance_loss,tensor,reference,numpy_reference,fp32,,numpy_reference,compile
 log_cosh_loss,tensor,reference,numpy_reference,fp32,,numpy_reference,compiled,,fp32,,manifest,tests/unit/test_apple_*.py,
 lookahead_sparse_attention,tensor,reference,numpy_reference,fp32,,numpy_reference,fused,,fp32,,manifest,tests/unit/test_apple_*.py,
 lora_linear,tensor,reference,numpy_reference,fp32,,numpy_reference,compiled,,fp32,,manifest,tests/unit/test_apple_*.py,
+lstm_cell,tensor,reference,numpy_reference,fp32,,numpy_reference,compiled,,fp32,,manifest,tests/unit/test_apple_*.py,
 lu,tensor,reference,numpy_reference,fp32,,numpy_reference,compiled,,fp32,,manifest,tests/unit/test_apple_*.py,
 mae_loss,tensor,reference,numpy_reference,fp32,,numpy_reference,compiled,,fp32,,manifest,tests/unit/test_apple_*.py,
 masked_categorical,tensor,reference,numpy_reference,fp32,,numpy_reference,fused,,"fp32,int32",tessera_apple_gpu_masked_categorical_f32,driver,tests/unit/test_apple_gpu_ldt_loss_ops.py,

diff --git a/docs/audit/generated/apple_target_map.md b/docs/audit/generated/apple_target_map.md
@@ -54,12 +54,12 @@ and ``docs/apple_backend.md`` (Metal 4 implementation-state review):
 
 | Family | Rows | apple_gpu fused | apple_cpu accelerate_native |
 |---|---:|---:|---:|
-| tensor | 134 | 21 | 2 |
+| tensor | 135 | 21 | 2 |
 | ga | 19 | 19 | 0 |
 | ebm | 14 | 14 | 0 |
 | m7 | 4 | 4 | 0 |
 
-## tensor (134)
+## tensor (135)
 
 | Op | apple_cpu status | cpu framework | cpu dtypes | cpu symbol | execution_kind | apple_gpu status | gpu framework | gpu dtypes | gpu symbol | gpu_dispatch | proof |
 |---|---|---|---|---|---|---|---|---|---|---|---|
@@ -125,6 +125,7 @@ and ``docs/apple_backend.md`` (Metal 4 implementation-state review):
 | log_cosh_loss | reference | numpy_reference | fp32 | - | numpy_reference | compiled | - | fp32 | - | manifest | tests/unit/test_apple_*.py |
 | lookahead_sparse_attention | reference | numpy_reference | fp32 | - | numpy_reference | fused | - | fp32 | - | manifest | tests/unit/test_apple_*.py |
 | lora_linear | reference | numpy_reference | fp32 | - | numpy_reference | compiled | - | fp32 | - | manifest | tests/unit/test_apple_*.py |
+| lstm_cell | reference | numpy_reference | fp32 | - | numpy_reference | compiled | - | fp32 | - | manifest | tests/unit/test_apple_*.py |
 | lu | reference | numpy_reference | fp32 | - | numpy_reference | compiled | - | fp32 | - | manifest | tests/unit/test_apple_*.py |
 | mae_loss | reference | numpy_reference | fp32 | - | numpy_reference | compiled | - | fp32 | - | manifest | tests/unit/test_apple_*.py |
 | masked_categorical | reference | numpy_reference | fp32 | - | numpy_reference | fused | - | fp32,int32 | tessera_apple_gpu_masked_categorical_f32 | driver | tests/unit/test_apple_gpu_ldt_loss_ops.py |

diff --git a/docs/audit/generated/compiler_progress.csv b/docs/audit/generated/compiler_progress.csv
@@ -15,13 +15,13 @@ primitive_axis,lowering_rule,closed,480,480,0,"primitive contract axis; open mea
 primitive_axis,backend_kernel,mixed,99,480,381,"primitive contract axis; open means partial or planned, not necessarily missing API support",docs/audit/generated/s_series_status.md,Promote by backend/pathway; do not treat every target as an all-up compiler veto.
 integration,Verifier coverage,mixed,163,174,11,"no_verifier=11, real=163",docs/audit/generated/verifier_coverage.csv,"Add real verifier implementations for no_verifier ops, prioritizing native codegen lanes."
 integration,Direct test evidence,mixed,363,480,117,"covered_by_family=26, directly_tested=363, hardware_gated=4, structural_only=87",docs/audit/generated/test_coverage.csv,Convert structural_only and needs_direct_test rows into direct compare fixtures; keep hardware_gated tied to backend proof.
-integration,Runtime execution matrix,closed,147,147,0,"apple_cpu=2, apple_gpu=16, cpu=2, nvidia_sm120=1, rocm=65, x86=61",docs/audit/generated/runtime_execution_matrix.csv,Add rows only when a launch path actually executes.
+integration,Runtime execution matrix,closed,149,149,0,"apple_cpu=2, apple_gpu=16, cpu=2, nvidia_sm120=1, rocm=67, x86=61",docs/audit/generated/runtime_execution_matrix.csv,Add rows only when a launch path actually executes.
 integration,Runtime ABI symbols,mixed,397,650,253,"apple=557, nvidia=7, rocm=10, x86=76",docs/audit/generated/runtime_abi.csv,Reduce stub-only ABI rows where a backend claims native execution.
 integration,Audited repo surfaces,mixed,31,58,27,"archived=4, compile_only=12, runnable=31, runnable_optional=1, scaffold=10",docs/audit/generated/surface_status.csv,Graduate compile_only/scaffold entries that exercise compiler pathways; archive dead surfaces.
-codegen_pathway,Apple CPU,closed,173,173,0,"runtime: apple_cpu=2; target_map: fused=2, reference=169",docs/audit/generated/apple_target_map.csv,Keep as regression baseline for CPU value-call/runtime ABI.
-codegen_pathway,Apple GPU,mixed,186,187,1,"runtime: apple_gpu=16; target_map: absent=1, compiled=101, fused=58, hardware_verified=11",docs/audit/generated/apple_target_map.csv,Close the remaining absent target-map lane or document why it is host-only.
+codegen_pathway,Apple CPU,closed,174,174,0,"runtime: apple_cpu=2; target_map: fused=2, reference=170",docs/audit/generated/apple_target_map.csv,Keep as regression baseline for CPU value-call/runtime ABI.
+codegen_pathway,Apple GPU,mixed,187,188,1,"runtime: apple_gpu=16; target_map: absent=1, compiled=102, fused=58, hardware_verified=11",docs/audit/generated/apple_target_map.csv,Close the remaining absent target-map lane or document why it is host-only.
 codegen_pathway,x86 / CPU,closed,2,2,0,runtime: cpu=2,docs/audit/generated/runtime_execution_matrix.csv,Keep native CPU and numpy reference lanes separate in runtime proofs.
-codegen_pathway,ROCm / HIP,closed,101,101,0,"runtime: rocm=65; target_map: compiled=34, hardware_verified=2",docs/audit/generated/rocm_target_map.csv,Close the artifact-only target-map tail and preserve CDNA as hardware-gated.
+codegen_pathway,ROCm / HIP,closed,103,103,0,"runtime: rocm=67; target_map: compiled=34, hardware_verified=2",docs/audit/generated/rocm_target_map.csv,Close the artifact-only target-map tail and preserve CDNA as hardware-gated.
 codegen_pathway,CUDA / NVIDIA,mixed,1,40,39,runtime: nvidia_sm120=1; target_map: artifact_only=39,docs/audit/generated/nvidia_sm90_target_map.csv,"Promote artifact-only rows with execute-and-compare, starting from sm_120 matmul adjacency and attention."
 open_work,Target IR native/fused codegen,mixed,309,315,6,"compiled=196, fused=71, hardware_verified=13, not_applicable=29, reference=6",docs/audit/generated/support_table.csv,Promote high-use reference rows into native/fused Target IR or mark intentional reference-only lanes.
 open_work,backend_kernel,mixed,99,480,381,"primitive contract axis; open means partial or planned, not necessarily missing API support",docs/audit/generated/s_series_status.md,Promote by backend/pathway; do not treat every target as an all-up compiler veto.

diff --git a/docs/audit/generated/compiler_progress.md b/docs/audit/generated/compiler_progress.md
@@ -42,18 +42,18 @@ A row is not marked incomplete merely because Apple, x86, ROCm, and CUDA are not
 |---|---|---:|---:|---:|---|---|
 | `Verifier coverage` | mixed | 163 | 174 | 11 | no_verifier=11, real=163 | Add real verifier implementations for no_verifier ops, prioritizing native codegen lanes. |
 | `Direct test evidence` | mixed | 363 | 480 | 117 | covered_by_family=26, directly_tested=363, hardware_gated=4, structural_only=87 | Convert structural_only and needs_direct_test rows into direct compare fixtures; keep hardware_gated tied to backend proof. |
-| `Runtime execution matrix` | closed | 147 | 147 | 0 | apple_cpu=2, apple_gpu=16, cpu=2, nvidia_sm120=1, rocm=65, x86=61 | Add rows only when a launch path actually executes. |
+| `Runtime execution matrix` | closed | 149 | 149 | 0 | apple_cpu=2, apple_gpu=16, cpu=2, nvidia_sm120=1, rocm=67, x86=61 | Add rows only when a launch path actually executes. |
 | `Runtime ABI symbols` | mixed | 397 | 650 | 253 | apple=557, nvidia=7, rocm=10, x86=76 | Reduce stub-only ABI rows where a backend claims native execution. |
 | `Audited repo surfaces` | mixed | 31 | 58 | 27 | archived=4, compile_only=12, runnable=31, runnable_optional=1, scaffold=10 | Graduate compile_only/scaffold entries that exercise compiler pathways; archive dead surfaces. |
 
 ## Code Generation Pathways
 
 | Item | Status | Ready | Total | Open | Detail | Next |
 |---|---|---:|---:|---:|---|---|
-| `Apple CPU` | closed | 173 | 173 | 0 | runtime: apple_cpu=2; target_map: fused=2, reference=169 | Keep as regression baseline for CPU value-call/runtime ABI. |
-| `Apple GPU` | mixed | 186 | 187 | 1 | runtime: apple_gpu=16; target_map: absent=1, compiled=101, fused=58, hardware_verified=11 | Close the remaining absent target-map lane or document why it is host-only. |
+| `Apple CPU` | closed | 174 | 174 | 0 | runtime: apple_cpu=2; target_map: fused=2, reference=170 | Keep as regression baseline for CPU value-call/runtime ABI. |
+| `Apple GPU` | mixed | 187 | 188 | 1 | runtime: apple_gpu=16; target_map: absent=1, compiled=102, fused=58, hardware_verified=11 | Close the remaining absent target-map lane or document why it is host-only. |
 | `x86 / CPU` | closed | 2 | 2 | 0 | runtime: cpu=2 | Keep native CPU and numpy reference lanes separate in runtime proofs. |
-| `ROCm / HIP` | closed | 101 | 101 | 0 | runtime: rocm=65; target_map: compiled=34, hardware_verified=2 | Close the artifact-only target-map tail and preserve CDNA as hardware-gated. |
+| `ROCm / HIP` | closed | 103 | 103 | 0 | runtime: rocm=67; target_map: compiled=34, hardware_verified=2 | Close the artifact-only target-map tail and preserve CDNA as hardware-gated. |
 | `CUDA / NVIDIA` | mixed | 1 | 40 | 39 | runtime: nvidia_sm120=1; target_map: artifact_only=39 | Promote artifact-only rows with execute-and-compare, starting from sm_120 matmul adjacency and attention. |
 
 ## Open Work Summary

diff --git a/docs/audit/generated/rocm_target_map.csv b/docs/audit/generated/rocm_target_map.csv
@@ -6,7 +6,7 @@ softmax,activation,compiled,"fp32,fp16,bf16",hipcc≥7.2.4,,,,"Row-wise stable s
 softmax_safe,activation,compiled,fp32,hipcc≥7.2.4,,,,Stable reduction softmax_safe — max-shifted reduce (max/sum) + unary exp/log lane. Executes via runtime.launch() (rocm_stable_reduce_compiled).
 attn_sliding_window,attention,compiled,"fp16,bf16",hipcc≥7.2.4,,,,"Mistral sliding-window via the flash_attn WMMA kernel (sliding_window attr, causal band of width W; KV-tile skip). Executes via runtime.launch() (window kwarg)."
 deepseek_sparse_attention,attention,compiled,fp32,hipcc≥7.2.4,,,,"DeepSeek/NSA composition — sliding + compressed-block branches remain reference compositions while the top-k branch uses the GPU-resident top-k selector plus selected-block sparse-attention kernel when ROCm is available. Executes via runtime.launch() (rocm_sparse_attn_compiled), with exact reference fallback off hardware."
-flash_attn,attention,hardware_verified,"fp16,bf16",hipcc≥7.2.4,,0.6500,,"RDNA 3.5 WMMA flash-attention forward executes on the AMD GPU through the shipped libtessera_rocm_flash_attn.so symbols (tessera_rocm_wmma_flash_attn_{f16,bf16}, HIPRTC-compiled for the device arch at load); ROCm 7.2.4. The second op after matmul to run natively on a non-Apple backend. Numerically validated vs a numpy attention reference by the execute_compare_fixture."
+flash_attn,attention,hardware_verified,"fp16,bf16",hipcc≥7.2.4,,0.6500,,"RDNA 3.5 WMMA flash-attention forward executes on the AMD GPU through the shipped libtessera_rocm_flash_attn.so symbols (tessera_rocm_wmma_flash_attn_{f16,bf16}, HIPRTC-compiled for the device arch at load); ROCm 7.2.4. The second op after matmul to run natively on a non-Apple backend. Numerically validated vs a numpy attention reference by the execute_compare_fixture. The FA-2 BACKWARD (dQ/dK/dV) also executes on gfx1151 via the compiler-generated rocm_flash_attn_bwd_compiled lane (generate-wmma-flash-attn-bwd-kernel -> fa_pre/fa_dkdv/fa_dq; MHA + GQA/MQA + additive attn_bias + sliding-window + logit-softcap, scale+causal), validated vs autodiff vjp_flash_attn — see the runtime_execution_matrix."
 gated_attention,attention,compiled,"fp16,bf16",hipcc≥7.2.4,,,,Softmax attention × a learned gate — the WMMA flash_attn kernel + an elementwise sigmoid-gate multiply. Executes via runtime.launch() (rocm_exotic_attn_compiled).
 gated_deltanet,attention,compiled,"fp32,fp16,bf16",hipcc≥7.2.4,,,,"Gated/delta linear-attention recurrence as a causal SEQUENTIAL-SCAN kernel (generate-rocm-deltanet-kernel: one workgroup per (b,h), one thread per value-column, LDS state) — the first recurrent compiled ROCm kernel. erase/gate/beta/decay flags. Executes via runtime.launch() (rocm_deltanet_compiled)."
 gqa_attention,attention,compiled,"fp16,bf16",hipcc≥7.2.4,,,,"GQA/MQA via the flash_attn WMMA kernel (gqa directive attr; fwd+bwd, grouped K/V; runtime detects from operand shapes). Executes on gfx1151 via runtime.launch() (rocm_flash_attn_compiled); no shipped C-ABI symbol."