score.sh: add 10-year position dimension + CI guard

New 6th scoring dimension explicitly measures the structural moats that
define our 10-year position. Previously the harness measured what we
built (structure/correctness/quality/perf/integration) but had no metric
for what makes us defensible vs llama.cpp.

Six new sub-dimensions (weights total 9):
  - single_header_loc       quant.h <= 16,000 lines (currently 15,785)
  - single_header_size      quant.h <= 700 KB        (currently 646 KB)
  - core_zero_deps          src/core/*.c only includes libc / SIMD intrinsics
                            / OS threading / project headers
  - papers_implemented      polar / qjl / turbo / uniform / turbo_kv = 5
  - honest_corrections      >=4 self-corrections logged in CHANGELOG
  - pypi_distribution       pyproject.toml + publish.yml present

Whitelisted in core_zero_deps: arm_neon.h, immintrin.h, wasm_simd128.h,
pthread.h, windows.h, sched.h, sys/* etc - these are all libc/OS, not
third-party deps.

Run history skip rule: --bench, --quality, --position no longer pollute
.score_history. Only --quick and --full write to the trend.

CI guard: .github/workflows/ci.yml runs `score.sh --position` on Linux
and fails the build if the position dimension drops below 75%. This
means any future change that bloats quant.h past 16K LOC, adds a
third-party dep to src/core, removes a paper implementation, etc.,
gets caught at PR time, not after PyPI ships.

Baseline (v0.8.1, Apple Silicon, --quick mode):
  total       96.1% (was 94.5% under 5-dim scoring)
  position   100.0%

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: unamedkr

.github/workflows/ci.yml

@@ -55,6 +55,28 @@ jobs:
       - name: Run tests
         run: ctest --test-dir build --output-on-failure --timeout 120 -C Release
 
+      # 10-year position guard: fail CI if structural moats erode.
+      # Runs on Linux only (POSIX shell). Checks single-header LOC/size,
+      # zero-deps, papers count, honest correction track, PyPI presence.
+      - name: 10-year position guard
+        if: matrix.os == 'ubuntu-latest'
+        shell: bash
+        run: |
+          chmod +x score.sh
+          bash score.sh --position 2>&1 | tee /tmp/position.log
+          # Extract dimension percentage; require >= 75%.
+          pct=$(grep -E '^\s*position\s' /tmp/position.log | awk '{print $2}' | tr -d '%')
+          echo "position dimension: ${pct}%"
+          if [ -z "$pct" ]; then
+            echo "could not parse position score"
+            exit 1
+          fi
+          # bash arithmetic: integer compare
+          if [ "${pct%.*}" -lt 75 ]; then
+            echo "::error::Position dimension regressed below 75% (${pct}%)"
+            exit 1
+          fi
+
       - name: Upload test results
         if: failure()
         uses: actions/upload-artifact@v4

score.sh

@@ -306,6 +306,103 @@ eval_integration() {
     print_item "documentation" "$dc" 3 1
 }
 
+# ============================================================
+# DIMENSION 6: 10-YEAR POSITION (structural moats)
+# ------------------------------------------------------------
+# These metrics protect the project's defensible position:
+# single-header embeddability, zero deps, research velocity,
+# claim audit-ability. Anything that erodes them should drag
+# the score, even if other dimensions look fine.
+# ============================================================
+eval_position() {
+    echo -e "\n${BOLD}${CYAN}[6/6] 10-YEAR POSITION (structural moats)${NC}"
+
+    # ----- Single-header LOC budget (≤ 16,000 lines) -----
+    local sh_loc=0
+    if [ -f "$PROJECT_DIR/quant.h" ]; then
+        sh_loc=$(wc -l < "$PROJECT_DIR/quant.h" | tr -d ' ')
+    fi
+    local sh_loc_score=0
+    if [ "$sh_loc" -gt 0 ] && [ "$sh_loc" -le 16000 ]; then
+        sh_loc_score=1
+    fi
+    log_score "position" "single_header_loc" "$sh_loc_score" 1 2
+    print_item "single_header_loc ($sh_loc / 16000)" "$sh_loc_score" 1 2
+
+    # ----- Single-header binary size budget (≤ 700 KB) -----
+    local sh_size=0
+    if [ -f "$PROJECT_DIR/quant.h" ]; then
+        # macOS / BSD stat -f%z, GNU stat -c%s — try both
+        sh_size=$(stat -f%z "$PROJECT_DIR/quant.h" 2>/dev/null || stat -c%s "$PROJECT_DIR/quant.h" 2>/dev/null || echo 0)
+    fi
+    local sh_size_kb=$((sh_size / 1024))
+    local sh_size_score=0
+    if [ "$sh_size_kb" -gt 0 ] && [ "$sh_size_kb" -le 700 ]; then
+        sh_size_score=1
+    fi
+    log_score "position" "single_header_size" "$sh_size_score" 1 1
+    print_item "single_header_size (${sh_size_kb} KB / 700)" "$sh_size_score" 1 1
+
+    # ----- Zero external dependencies in core (libc/libm/intrinsics/OS) -----
+    # Allowed:
+    #   - C standard library headers
+    #   - SIMD intrinsics (arm_neon.h, immintrin.h, wasm_simd128.h)
+    #   - OS threading / kernel headers (pthread.h, windows.h, sched.h)
+    #   - Project headers (turboquant/*, tq_*)
+    # A failure here means we picked up a real third-party dep.
+    local bad_includes=0
+    if [ -d "$PROJECT_DIR/src/core" ]; then
+        bad_includes=$(grep -hE '^[[:space:]]*#include[[:space:]]*[<"]' "$PROJECT_DIR/src/core/"*.c 2>/dev/null \
+            | grep -vE '<(stdint|string|math|stdlib|stdio|stddef|stdbool|assert|float|limits|inttypes|errno|time|ctype|signal)\.h>' \
+            | grep -vE '<(arm_neon|immintrin|wasm_simd128|x86intrin|emmintrin|smmintrin|tmmintrin|nmmintrin|avxintrin|avx2intrin)\.h>' \
+            | grep -vE '<(pthread|sched|unistd|sys/[a-z_]+|windows|fcntl)\.h>' \
+            | grep -vE '"(turboquant/|tq_)' \
+            | wc -l | tr -d ' ')
+    fi
+    local deps_score=0
+    [ "$bad_includes" = "0" ] && deps_score=1
+    log_score "position" "core_zero_deps" "$deps_score" 1 2
+    print_item "core_zero_deps ($bad_includes foreign includes)" "$deps_score" 1 2
+
+    # ----- Papers ported (research velocity proxy) -----
+    # Counts implementation files matching known KV-quant paper algorithms.
+    # Goal is +1 every quarter; baseline as of v0.8.0 = 5 (polar, qjl, turbo,
+    # uniform, turbo_kv). Score reflects whether we're maintaining the count.
+    local papers=0
+    [ -f "$PROJECT_DIR/src/core/tq_polar.c" ]    && papers=$((papers + 1))
+    [ -f "$PROJECT_DIR/src/core/tq_qjl.c" ]      && papers=$((papers + 1))
+    [ -f "$PROJECT_DIR/src/core/tq_turbo.c" ]    && papers=$((papers + 1))
+    [ -f "$PROJECT_DIR/src/core/tq_uniform.c" ]  && papers=$((papers + 1))
+    [ -f "$PROJECT_DIR/src/core/tq_turbo_kv.c" ] && papers=$((papers + 1))
+    log_score "position" "papers_implemented" "$papers" 5 2
+    print_item "papers_implemented" "$papers" 5 2
+
+    # ----- Honest correction track (CHANGELOG retrospective entries) -----
+    # Counts CHANGELOG headings that name a self-correction. Reframes
+    # corrections as a positive — they're our trust asset.
+    local corrections=0
+    if [ -f "$PROJECT_DIR/CHANGELOG.md" ]; then
+        corrections=$(grep -ciE 'honest correction|self.?corrected|hotfix|retracted|retract' \
+                      "$PROJECT_DIR/CHANGELOG.md" 2>/dev/null || echo 0)
+    fi
+    # Cap at 10 — beyond that the metric stops rewarding new ones.
+    [ "$corrections" -gt 10 ] && corrections=10
+    local correction_score=0
+    [ "$corrections" -ge 4 ] && correction_score=1
+    log_score "position" "honest_corrections" "$correction_score" 1 1
+    print_item "honest_corrections ($corrections logged)" "$correction_score" 1 1
+
+    # ----- PyPI distribution channel live -----
+    local pypi_live=0
+    if [ -f "$PROJECT_DIR/bindings/python/pyproject.toml" ] && \
+       grep -q '^name *= *"quantcpp"' "$PROJECT_DIR/bindings/python/pyproject.toml" && \
+       [ -f "$PROJECT_DIR/.github/workflows/publish.yml" ]; then
+        pypi_live=1
+    fi
+    log_score "position" "pypi_distribution" "$pypi_live" 1 1
+    print_item "pypi_distribution" "$pypi_live" 1 1
+}
+
 # ============================================================
 # FINAL REPORT
 # ============================================================
@@ -333,7 +430,7 @@ print_final() {
     # Dimension breakdown
     echo ""
     echo "  Dimension Breakdown:"
-    for dim in structure correctness quality performance integration; do
+    for dim in structure correctness quality performance integration position; do
         local ds=0 dw=0
         while IFS='|' read -r cat name score max weight; do
             if [ "$cat" = "$dim" ]; then
@@ -350,8 +447,16 @@ print_final() {
         fi
     done
 
-    # Save history
-    echo "$(date '+%Y-%m-%d %H:%M:%S') $final" >> "$SCORE_LOG"
+    # Save history — only for full / quick evaluations.
+    # Single-dimension modes (--bench, --quality, --position) skip the log
+    # so partial scores don't pollute the trend line.
+    case "$MODE" in
+        --bench|--quality|--position)
+            ;;
+        *)
+            echo "$(date '+%Y-%m-%d %H:%M:%S') $final" >> "$SCORE_LOG"
+            ;;
+    esac
 
     # Trend
     if [ -f "$SCORE_LOG" ] && [ "$(wc -l < "$SCORE_LOG" | tr -d ' ')" -gt 1 ]; then
@@ -379,19 +484,24 @@ case "$MODE" in
     --quick)
         eval_structure
         eval_correctness
+        eval_position
         ;;
     --bench)
         eval_performance
         ;;
     --quality)
         eval_quality
         ;;
+    --position)
+        eval_position
+        ;;
     --full|*)
         eval_structure
         eval_correctness
         eval_quality
         eval_performance
         eval_integration
+        eval_position
         ;;
 esac