feat(power): per-worker power + per-stage J/token for disagg

Extends multinode measured-power aggregation with per-worker breakdown
and per-stage joules attribution. The cluster-wide avg_power_w +
joules_per_*_token fields stay backward-compatible; new disagg-only
fields layer on top.

New agg JSON fields:
  - power_by_worker: list of {role, worker_idx, hosts, num_gpus,
    avg_power_w} parsed from srt-slurm perfmon CSV filenames
    (`perf_samples_<role>_w<idx>_<host>.csv`). Roles: prefill, decode,
    agg, frontend. Workers spanning N nodes collapse one entry whose
    num_gpus is the cross-node sum.
  - joules_per_input_token: prefill_energy / total_input_tokens
    (disagg only — meaningless without a prefill stage).

Per-stage attribution (disagg only) replaces cluster-wide ratios for
existing fields:
  - joules_per_output_token = decode_energy / output_tokens
  - joules_per_total_token = (prefill + decode) / all_tokens
Frontend-labeled CSVs are excluded from per-stage energy but still
listed for observability. Falls back to cluster-wide math if only one
stage's CSVs survived.

process_result.py now passes DISAGG through to aggregate_power.run().
launch_gb300-cw.sh's recipe-injection loop reports found/injected
counts so a zero-recipes-found bug is distinguishable from the
benign all-already-monitored case.

Tests: 88/88 pass (68 existing + 20 new). New coverage: label parsing
across host formats, multi-node-per-worker collapse, per-stage J/token
math, frontend exclusion, single-stage fallback, zero-input degenerate,
end-to-end disagg wiring through process_result.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Author: arygupt

runners/launch_gb300-cw.sh

@@ -123,16 +123,25 @@ cp -rT "$SRT_RECIPE_SRC" "$SRT_RECIPE_DST"
 # and perfmon never spawns. PR #1574's first real sweep (#26548110246) hit
 # exactly this: completed "success" with no power data because the glob
 # matched nothing and the failure was silent end-to-end.
+FOUND_COUNT=0
 INJECTED_COUNT=0
 while IFS= read -r recipe; do
+    FOUND_COUNT=$((FOUND_COUNT + 1))
     if ! grep -q '^monitoring:' "$recipe"; then
         printf '\nmonitoring:\n  enabled: true\n  sample_interval: 1.0\n' >> "$recipe"
         echo "[perfmon] enabled monitoring in recipe: $recipe"
         INJECTED_COUNT=$((INJECTED_COUNT + 1))
     fi
 done < <(find "$SRT_RECIPE_DST" -type f -name '*.yaml')
-if [ "$INJECTED_COUNT" -eq 0 ]; then
-    echo "[perfmon] WARNING: zero recipes received monitoring injection under $SRT_RECIPE_DST. Either every recipe already had it, or the directory layout changed — power data will be MISSING from this run." >&2
+# Distinguish "found 0 recipes" (real bug — directory wrong/empty) from "all
+# already had monitoring:" (benign — happens on reruns or if a recipe author
+# pre-declared the block). Only the former is a missing-power-data risk.
+if [ "$FOUND_COUNT" -eq 0 ]; then
+    echo "[perfmon] WARNING: zero recipe YAMLs found under $SRT_RECIPE_DST. The directory layout may have changed — power data will be MISSING from this run." >&2
+elif [ "$INJECTED_COUNT" -eq 0 ]; then
+    echo "[perfmon] all $FOUND_COUNT recipes already declared monitoring: — no injection needed."
+else
+    echo "[perfmon] injected monitoring: into $INJECTED_COUNT of $FOUND_COUNT recipes."
 fi
 
 echo "Installing srtctl..."