#!/bin/sh
set -eu

[ "$#" -eq 2 ] || [ "$#" -eq 3 ] || {
    echo "usage: $0 BASELINE_JSON CANDIDATE_JSON [OUTPUT_JSON]" >&2
    exit 2
}
baseline=$1
candidate=$2
output=${3:-}

analyze() {
    jq -n --slurpfile baseline "$baseline" --slurpfile candidate "$candidate" '
      def mean($xs): $xs | add / length;
      def sample_stddev($xs):
        mean($xs) as $m |
        (($xs | map((. - $m) * (. - $m)) | add) / (($xs | length) - 1)) | sqrt;
      def cv_pct($xs): 100 * sample_stddev($xs) / mean($xs);
      def improvement_pct($before; $after): 100 * (1 - mean($after) / mean($before));

      $baseline[0] as $b | $candidate[0] as $c |
      [
        "schema_version", "engine", "workload", "shape", "query_heads", "kv_heads",
        "head_dimension", "tokens", "live_tokens", "cache_capacity", "window",
        "kv_cache_bytes", "score_scratch_bytes", "value_kernel",
        "value_dimension_tile", "value_threadgroups", "value_partition_scratch_bytes",
        "value_partitions", "iterations", "repetitions", "warmup", "dispatches"
      ] as $controls |
      ($controls | map(select($b[.] != $c[.]))) as $changed_controls |
      if ($b.schema_version != 1 or $c.schema_version != 1 or
          $b.workload != "synthetic-exact-shape-decode-attention-stage-profile" or
          $c.workload != $b.workload or
          $changed_controls != [] or
          ($b.score_token_tile | type) != "number" or
          ($c.score_token_tile | type) != "number" or
          $b.score_token_tile < 1 or $c.score_token_tile < 1 or
          $b.score_token_tile == $c.score_token_tile or
          ($b.gpu_seconds_samples | length) != $b.repetitions or
          ($c.gpu_seconds_samples | length) != $c.repetitions or
          ($b.stages.score.samples_seconds | length) != $b.repetitions or
          ($c.stages.score.samples_seconds | length) != $c.repetitions or
          $b.repetitions < 3 or $c.repetitions < 3)
      then error("decode-attention inputs are not a valid one-axis score-token-tile comparison")
      else
        improvement_pct($b.gpu_seconds_samples; $c.gpu_seconds_samples) as $total_gain |
        improvement_pct($b.stages.score.samples_seconds; $c.stages.score.samples_seconds) as $score_gain |
        cv_pct($b.gpu_seconds_samples) as $bt_cv |
        cv_pct($c.gpu_seconds_samples) as $ct_cv |
        cv_pct($b.stages.score.samples_seconds) as $bs_cv |
        cv_pct($c.stages.score.samples_seconds) as $cs_cv |
        ([$bt_cv, $ct_cv, $bs_cv, $cs_cv] | max) as $noise |
        {
          schema_version: 1,
          scope: "synthetic exact-shape decode-attention microbenchmark",
          axis: {
            name: "score_token_tile",
            baseline: $b.score_token_tile,
            candidate: $c.score_token_tile
          },
          controls: ($controls | map({key: ., value: $b[.]} ) | from_entries),
          total_gpu: {
            baseline_mean_seconds: mean($b.gpu_seconds_samples),
            candidate_mean_seconds: mean($c.gpu_seconds_samples),
            improvement_pct: $total_gain,
            baseline_cv_pct: $bt_cv,
            candidate_cv_pct: $ct_cv
          },
          score_stage: {
            baseline_mean_seconds: mean($b.stages.score.samples_seconds),
            candidate_mean_seconds: mean($c.stages.score.samples_seconds),
            improvement_pct: $score_gain,
            baseline_cv_pct: $bs_cv,
            candidate_cv_pct: $cs_cv
          },
          observed_noise_floor_pct: $noise,
          screening: {
            threshold_passed: ($total_gain >= 3 and $total_gain > $noise),
            promotion_eligible: false,
            reason: "microbenchmark only; native correctness and authentic populated-model gates remain required"
          }
        }
      end'
}

if [ -n "$output" ]; then analyze > "$output"; else analyze; fi