feat(tooling): run-atlas-naming.sh — overnight batch naming launcher (#833)

Wraps gemma_naming.py with the validated overnight recipe: gfx1201
ROCm binary path, distrobox reach-build for libhipblas at runtime,
timestamped log under .tmp/.

Preflight checks: binary exists and is executable, model present,
reach-build container exists, binary strings contains gfx1201 kernels.
Fails fast on any missing prerequisite so a broken build can't waste
an overnight window. Script takes no arguments; anything passed is
rejected so a stray --help can't accidentally launch the pipeline.

Estimate ~4-6 h for ~26k features across 2394 bodies at 74 t/s on an
RX 9070. Safe to interrupt and resume — preserved path skips
already-named bodies.

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

tooling/planet-gen/run-atlas-naming.sh

@@ -0,0 +1,77 @@
+#!/usr/bin/env bash
+# Run the full Gemma 2 batch naming pipeline across every markers.json
+# in wiki/star-systems. Uses the gfx1201 ROCm binary via distrobox.
+#
+# Safe to interrupt and resume: the pipeline preserves bodies that
+# already have non-empty name fields, so re-running picks up where it
+# left off. A crash loses at most the current body's in-progress work.
+#
+# Wall time on an RX 9070 is roughly 4-6 h for the ~26k features in
+# 2394 bodies. Run with `nohup` and tail the log file:
+#
+#   nohup tooling/planet-gen/run-atlas-naming.sh > /tmp/atlas.out 2>&1 &
+#   tail -f .tmp/atlas-naming-*.log
+
+set -euo pipefail
+
+# No arguments accepted — everything is hardcoded for the overnight
+# run. Guard against typos/--help slipping through to gemma_naming.py
+# and starting a real run when the caller expected a help screen.
+if [[ $# -gt 0 ]]; then
+    echo "usage: $0" >&2
+    echo "  (no arguments; edit this script to change binary/model/container)" >&2
+    exit 2
+fi
+
+REPO_ROOT="$(cd "$(dirname "$0")/../.." && pwd)"
+cd "$REPO_ROOT"
+
+LOG_DIR="$REPO_ROOT/.tmp"
+mkdir -p "$LOG_DIR"
+STAMP="$(date +%Y%m%d-%H%M%S)"
+LOG="$LOG_DIR/atlas-naming-$STAMP.log"
+
+BIN="$HOME/Projects/settled-reach/binaries/sr-voice-rocm"
+MODEL="/var/mnt/data/projects/settled-reach/main/server/models/gemma2.gguf"
+DISTROBOX_NAME="reach-build"
+
+if [[ ! -x "$BIN" ]]; then
+    echo "error: sr-voice binary not found at $BIN" >&2
+    echo "       build with --features rocm inside $DISTROBOX_NAME" >&2
+    exit 1
+fi
+if [[ ! -f "$MODEL" ]]; then
+    echo "error: Gemma 2 model not found at $MODEL" >&2
+    exit 1
+fi
+if ! distrobox list 2>/dev/null | grep -q "^[[:xdigit:]]\+ *| *$DISTROBOX_NAME "; then
+    echo "error: distrobox container '$DISTROBOX_NAME' not found" >&2
+    exit 1
+fi
+
+# Sanity-check the binary is compiled for the host GPU. The llama.cpp
+# HIP kernels embed their target architecture as substrings like
+# "amdgcn-amd-amdhsa--gfx1201". If gfx1201 is missing and e.g. only
+# gfx906 is present, the build shipped kernels for a different arch
+# and every inference will crash with "invalid device function".
+if ! strings "$BIN" 2>/dev/null | grep -qw gfx1201; then
+    echo "error: $BIN does not contain gfx1201 kernels" >&2
+    echo "       expected target for AMD Radeon RX 9070 (Navi 48)" >&2
+    echo "       rebuild with CMAKE_HIP_ARCHITECTURES=gfx1201 inside $DISTROBOX_NAME" >&2
+    exit 1
+fi
+
+echo "atlas naming run starting"
+echo "  bin:       $BIN"
+echo "  model:     $MODEL"
+echo "  distrobox: $DISTROBOX_NAME"
+echo "  log:       $LOG"
+echo "  bodies:    ~2394 (resume-safe — already-named skipped)"
+echo "  estimate:  ~4-6 h on an RX 9070"
+echo
+
+exec python3 tooling/planet-gen/gemma_naming.py \
+    --sr-voice "$BIN" \
+    --model "$MODEL" \
+    --distrobox "$DISTROBOX_NAME" \
+    --log "$LOG"