Eliminate repeated invariant work during ITS metrics collection and export #4453

Pre-filing checklist

• I searched existing issues and didn't find a duplicate

Component(s)

Rust OTAP dataflow (rust/otap-dataflow/)

Objective

Eliminate repeated work on invariant metadata during ITS metrics SDK collection
and export. This concerns the native ITS metrics implementation, not the removed
Rust OpenTelemetry metrics SDK.

Treat this as a category of avoidable collection/export overhead, starting with
two confirmed examples: repeated view resolution and repeated scope-attribute
encoding. Share immutable metadata and precomputed results where appropriate,
rather than rediscovering them on every reporting interval.

Rationale

1. View selection is recomputed for every export.

MetricsOtlpEncoder::encode() prepares each export batch. When views are
configured, prepare_metric_set_with_views() matches scope selectors once per
metric set in that batch, then resolves instrument selectors for each metric
field. These results are not retained across exports, even though the registered
entity attributes, metric descriptors, and the encoder's view configuration are
unchanged.

This is not measurement-recording overhead, and the no-views path skips it.
However, the same selection work repeats on every collection/export interval and
scales with metric sets, fields, and views.

2. Metrics repeatedly serialize immutable scope attributes.

Registered entities snapshot their typed attributes into EntityAttributeSet.
Metric exports retain those attributes through an Arc, but
metrics::otlp::encode_scope() iterates and protobuf-encodes them again for
every emitted scope.

The logging path already has a separate ScopeToBytesMap for encoded
scope-attribute fragments. Both signals use OTLP
InstrumentationScope.attributes; immutable attribute fragments are candidates
for shared pre-encoding, rather than separate serialization lifecycles.

Scope

• Resolve view selections and invariant output metadata once per metric-set
  identity and active view configuration. Registration is one possible boundary;
  an encoder-owned first-use plan is another where view configuration is not
  available at registration. Keep genuinely batch-dependent work separate.
• Pre-encode complete, validated entity scope-attribute fragments on a cold path
  and reuse them for both metrics and logs. Retain typed attributes for
  selection, equality, and deduplication. Scope names and outer envelopes remain
  signal-specific.
• Account for entity registration, deduplication, unregistration, pending
  exports, and replacement of view configuration. Reuse must not introduce stale
  identity matches, unbounded cache growth, or new shared locks in measurement
  recording.
• Review adjacent recurring metadata validation, projection, and collision-index
  construction for invariant portions that can be moved into the same plan. Do
  not assume all batch-dependent validation can be removed.
• Resolve semantic differences before sharing an encoder or fragment. For
  example, metrics currently clamp UInt to i64::MAX, while self-tracing
  casts it to i64. Choose and test the intended behavior explicitly.
• Keep invalid scope metadata explicit: cold-path validation/encoding must not
  silently accept truncation or overflow. Avoid carrying newly introduced
  truncation bookkeeping through every export merely to enable implementation
  reuse.

Acceptance Criteria

• Repeated exports with unchanged metadata do not repeat scope/instrument
  view matching.
• Repeated exports do not re-encode unchanged entity scope attributes, and
  metrics/logs can reuse the same validated attribute fragment.
• Tests preserve typed selector semantics, view fan-out and naming,
  collision detection, and OTLP output; cover entity and view-configuration
  lifecycle changes.
• Tests demonstrate bounded metadata retention and correct behavior for
  invalid attributes and identity reuse.
• Extend internal_metrics_bridge measurements to distinguish cold
  initialization/first export from steady-state repeated exports, with and
  without views at representative entity counts.
• Report before/after time, allocations, and retained memory, and verify
  that sharing scope encoding does not regress existing self-tracing benchmarks.

Dependencies or Blockers

• Scope attributes are pre-calculated but not limited or recycled #1904 tracks unbounded scope-cache retention and
  recycling; coordinate lifecycle design with that work.
• Expand MetricSelector to support scope_attribute view selectors #2742 added scope-attribute view selectors but did
  not address repeated matching.
• Avoid Prost in the Metrics ITS path #3668 removed Prost from the ITS metrics path but did
  not eliminate repeated invariant work.

Additional Context

Source references at upstream revision f3215c1c0:

• Export preparation and no-views fast path
• Per-export view resolution and typed scope matching
• Repeated scope-attribute encoding
• Registered attribute snapshot
• Existing internal metrics bridge benchmark

Historical measurements in #3714 reported 98.03 us
without views versus 179.59 us with six views for 100 entities. This includes
all view processing, not just matching; it motivates measuring the invariant
work separately, not attributing the entire difference to these two issues.