{
  "schemaVersion": 1,
  "backend": "metal",
  "sourceOfTruth": "Runtime-ready means a built llama.cpp fork graph route selects the shipped Metal kernel and a numeric smoke test passes. Symbol presence in default.metallib is not enough.",
  "latestReport": "reports/porting/2026-05-11/remaining-work-ledger.md",
  "kernels": {
    "turbo3": {
      "runtimeCapabilityKey": "turbo3",
      "status": "runtime-ready",
      "runtimeReady": true,
      "shippedSymbols": [
        "kernel_turbo3_dot",
        "kernel_turbo3_dot_multi"
      ],
      "graphOp": "GGML_OP_ATTN_SCORE_TBQ",
      "smokeTarget": "dispatch-smoke",
      "smokeCommand": "make -C packages/inference/verify dispatch-smoke",
      "smokeScores": 32,
      "maxDiff": 4.768e-7,
      "evidenceDate": "2026-05-11",
      "notes": "dispatch_smoke links against the built fork's libggml-metal.dylib and drives GGML_OP_ATTN_SCORE_TBQ through kernel_turbo3_dot_multi."
    },
    "turbo4": {
      "runtimeCapabilityKey": "turbo4",
      "status": "runtime-ready",
      "runtimeReady": true,
      "shippedSymbols": [
        "kernel_turbo4_dot",
        "kernel_turbo4_dot_multi"
      ],
      "graphOp": "GGML_OP_ATTN_SCORE_TBQ",
      "smokeTarget": "dispatch-smoke",
      "smokeCommand": "make -C packages/inference/verify dispatch-smoke",
      "smokeScores": 32,
      "maxDiff": 4.768e-7,
      "evidenceDate": "2026-05-11",
      "notes": "kernel_turbo4_dot(_multi) consumes the fork GGML_TYPE_TBQ4_0 layout: four 32-wide block_tbq4_0 records per 128-row (ggml_row_size=72). dispatch_smoke drives GGML_OP_ATTN_SCORE_TBQ through kernel_turbo4_dot_multi and numerically checks the reference dot."
    },
    "turbo3_tcq": {
      "runtimeCapabilityKey": "turbo3_tcq",
      "status": "runtime-ready",
      "runtimeReady": true,
      "shippedSymbols": [
        "kernel_turbo3_tcq_dot",
        "kernel_turbo3_tcq_dot_multi"
      ],
      "graphOp": "GGML_OP_ATTN_SCORE_TBQ",
      "smokeTarget": "dispatch-smoke",
      "smokeCommand": "make -C packages/inference/verify dispatch-smoke",
      "smokeScores": 32,
      "maxDiff": 4.768e-7,
      "evidenceDate": "2026-05-11",
      "notes": "The patch adds GGML_TYPE_TBQ3_TCQ row-size traits and drives GGML_OP_ATTN_SCORE_TBQ through kernel_turbo3_tcq_dot_multi with the TCQ codebook bound."
    },
    "qjl": {
      "runtimeCapabilityKey": "qjl_full",
      "status": "runtime-ready",
      "runtimeReady": true,
      "shippedSymbols": [
        "kernel_attn_score_qjl1_256",
        "kernel_attn_score_qjl1_256_multi"
      ],
      "graphOp": "GGML_OP_ATTN_SCORE_QJL",
      "smokeTarget": "dispatch-smoke",
      "smokeCommand": "make -C packages/inference/verify dispatch-smoke",
      "smokeScores": 32,
      "maxDiff": 2.384e-7,
      "evidenceDate": "2026-05-11",
      "notes": "dispatch_smoke links against the built fork's libggml-metal.dylib and drives GGML_OP_ATTN_SCORE_QJL through kernel_attn_score_qjl1_256_multi. Runtime tokens_per_threadgroup remains default 32, with ELIZA_METAL_QJL_TOKENS_PER_TG available for per-device autotune experiments."
    },
    "polar": {
      "runtimeCapabilityKey": "polarquant",
      "status": "runtime-ready",
      "runtimeReady": true,
      "shippedSymbols": [
        "kernel_get_rows_q4_polar",
        "kernel_mul_mv_q4_polar_f32",
        "kernel_mul_mv_q4_polar_preht_f32",
        "kernel_attn_score_q4_polar_preht_f32"
      ],
      "graphOp": "GGML_OP_ATTN_SCORE_POLAR",
      "smokeTarget": "dispatch-smoke",
      "smokeCommand": "make -C packages/inference/verify dispatch-smoke",
      "smokeScores": 128,
      "maxDiff": 3.815e-6,
      "evidenceDate": "2026-05-11",
      "notes": "dispatch_smoke covers raw-q use_qjl=0/1 through kernel_mul_mv_q4_polar_f32 and pre-Hadamard use_qjl=0/1 through ggml_attn_score_polar_preht plus kernel_attn_score_q4_polar_preht_f32. The preHT route is only valid when the graph explicitly supplies H*q."
    }
  }
}