Revert agent-side stats.max/avg export in favor of computing max/avg
via PromQL window expressions on the manager. Container live-stat
fan-out grows from 3 to 5 queries (gauge / diff / rate / max / avg).

- Re-introduce MAX / AVG to ValueType plus from_legacy_live_stat_label
  to map "stats.max" / "stats.avg" labels back into typed value types.
- Build max/avg templates that union a gauge sub-expression with a
  rate sub-expression and label_replace the result back to the legacy
  "stats.max" / "stats.avg" label.
- Classify metrics into gauge-shape vs rate-shape stats sources, with
  exact names for built-ins and regex patterns for accelerator metrics.
- Repository merges 5 query responses in a generic loop instead of
  unpacking three buckets.

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

…s_query

The gauge and rate branches both produced the same
label_replace + *_over_time + sum_by skeleton, differing only in
whether the inner selector was wrapped with rate(...[window]). The
two branches embedded that skeleton as inline f-strings, which
hid the symmetry and was hard to read.

Extract two methods on FixedQueryBuilder:

- _utilization_selector() — builds {METRIC}{labels}, hiding the
  brace-doubling away from rendering callers.
- _window_stat_subquery() — wraps a selector in
  label_replace({stat_fn}((sum by ({group_by})({selector}))[window:]),
  "value_type","{label}","value_type",".*"). This is the single
  place where the stats-subquery shape lives.

_render_stats_query now reads as: pick a regex, build a selector
(optionally wrapped in rate()), and feed it through the same
skeleton. Rendered output is byte-for-byte unchanged
(characterization tests pass without touching the fixture).

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

…ture

The window-stat gauge patterns lagged behind what legacy accelerator
plugins actually publish:

- *_power: legacy emits both stats.max and stats.avg; this PR was
  emitting only stats.max.
- *_temperature: legacy emits both stats.max and stats.avg; this PR
  was emitting neither.

Surveyed plugins (rebellions/common, rebellions/atom_max, habana, ipu,
mock):
- All emit *_mem (max only) and *_util (max + avg).
- Only mock currently emits *_power and *_temperature, both with
  {avg, max} filters.

Extend STATS_MAX_GAUGE_METRIC_PATTERNS to include _temperature and
STATS_AVG_GAUGE_METRIC_PATTERNS to include _power and _temperature so
the new pipeline matches what every legacy plugin actually publishes.

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

…bels

Earlier the new pipeline emitted "stats.max" / "stats.avg" / "stats.rate"
as the rewritten value_type label, mirroring the legacy live_stat dict
keys. That coupling was unnecessary: the labels are synthesized at
query time via label_replace and never persist in the Prometheus TSDB
(only "current" / "capacity" series actually get scraped). The legacy
"stats." prefix is a Valkey live_stat artifact that future legacy-
compat consumers should re-attach in a dedicated converter, not
something the manager pipeline needs to carry.

The new pipeline now emits clean, ValueType-enum-aligned labels:

    stats.max  -> max
    stats.avg  -> avg
    stats.rate -> rate

Round-trip helpers go away:

- ValueType.to_legacy_live_stat_label() removed (was a 1-arm match
  that always returned "stats.{value}" for the three stat slots).
- ValueType.from_live_stat_label() removed; KernelMetricValuesByKernel
  parses the response label via ValueType(value_type_str) directly.

Drop trailing .value on f-string interpolation of ValueType members —
StrEnum's __format__ already returns the string value, so
f"{ValueType.MAX}" == "max" without the explicit attribute access.

Test fixtures and the from_prometheus_response characterisation test
now use the bare labels.

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

The original phrasing led with the implementation ("window-based max/avg
queries via PromQL") rather than the user-facing motivation ("a
parallel supply for legacy stats.* live_stat fields that doesn't suffer
the agent-restart / accumulator failure modes"). Updates the changelog
entry to lead with the compat angle, expands coverage to include
stats.rate (now also produced), and notes the restart-safe / window
semantics that make this a meaningful upgrade over the legacy producer.

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

…oint

Derive STATS_MAX_GAUGE_METRIC_PATTERNS / STATS_AVG_GAUGE_METRIC_PATTERNS
from a single _ACCEL_GAUGE_SUFFIXES_* source so adding a new accelerator
metric kind (e.g. clock, voltage) is a one-suffix edit instead of editing
two regex strings in lockstep.

Functionally identical — the generated regex is the same alternation, just
sorted alphabetically. Snapshot tests adjusted for the new sort order.

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