[metric] feat: add weighted aggregation to Metric/reduce_metrics (#1508) - #7501

…-project#1508)

Implements the two independent proposals from the verl-project#1508 issue thread
(2026-07-06 and 2026-07-24 comments): PR verl-project#4778 introduced `Metric` and
`AggregationType` but did not carry a per-value weight through
aggregation, so DP ranks or micro-batches with unequal sample/token
counts were still averaged as if equal.

- `Metric.__init__`/`append`/`extend` now optionally accept a `weight`
  per value (stored as a parallel `self.weights` list, defaulting each
  value to 1.0 when omitted). A scalar weight passed to `extend`
  broadcasts to every value being added.
- `Metric.aggregate()` / `Metric._aggregate()`: MEAN uses
  `np.average(values, weights=weights)` and SUM uses
  `sum(value * weight)` whenever any weight differs from 1.0; when all
  weights are 1.0 (or omitted) the exact original `np.mean`/`np.sum`
  code path runs unchanged, so behavior is byte-identical for existing
  callers.
- `Metric.aggregate_dp()` now accepts an optional per-rank `weights`
  list (e.g. each rank's local sample/token count). MEAN/SUM combine
  ranks via `np.average(value_arrays, axis=0, weights=weights)`
  (Σ(w·v)/Σw) instead of a plain `np.mean` when weights are supplied;
  MIN/MAX are unaffected (still flatten-then-reduce). Omitting
  `weights` reproduces the previous unweighted behavior exactly.
- `reduce_metrics`: fixed the substring-match bug also raised on the
  issue thread, where `"min"/"max" in key` misfires on keys like
  "...minibatch..." or "...maximum...". Keys are now split on
  non-alphanumeric characters and "min"/"max" must appear as a
  standalone token (e.g. "loss/min", "loss_min") to trigger that
  reduction. Audited every metric key currently produced across
  verl/recipe (compute_data_metrics, losses.py, rollout_corr_helper.py,
  detach_utils.py, etc.) and none rely on the old substring-only
  behavior, so this tightening is safe.

Backward compatible: default weight is 1.0 everywhere, and
`aggregate()`/`aggregate_dp()` fall back to the exact unweighted
numpy calls when no non-default weights are present (verified
numerically in the new tests, e.g. test_weighted_mean_with_unit_weights_
matches_plain_mean, test_weighted_sum_with_unit_weights_matches_plain_
sum, test_aggregate_dp_unweighted_still_matches_plain_mean).

Adds CPU-only unit tests to tests/trainer/ppo/test_metric_utils_on_cpu.py
covering: weighted Metric.aggregate() vs hand-computed weighted mean,
unit-weight byte-for-byte backward compatibility, weighted SUM,
append/extend weight plumbing (including scalar-weight broadcast and
length-mismatch errors), weighted aggregate_dp across DP ranks with
unequal sample counts (asserting against the ground-truth mean of all
flattened raw samples, where the naive unweighted mean-of-means is
visibly wrong), and the reduce_metrics min/max substring-token fix.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>