Trace import evidence and limits

Span-based import emits capture-wide import metadata and operation-level import.latency and import.attributes mappings. The parsed synth.Config and OperationConfig retain these mappings as opaque informational values, including unknown evidence fields. They are not validated as execution fields or copied into a built topology or emitted attributes. A valid topology remains runnable with insufficient, unfamiliar, or user-edited evidence.

Own time

For each effective operation invocation, own time is elapsed span duration minus the union of its effective downstream children's intervals. Intervals are clipped to the parent's bounds; overlapping time is subtracted once. Negative elapsed times become zero estimates. The existing emitter requires a positive mean, so the rounded emitted mean has a 1 microsecond floor.

Same-operation continuation spans are folded by the existing call inference; their downstream calls attach to the enclosing operation. Latency observations refer to effective invocations; attribute observations include every source span of that operation, including continuations.

An out-of-bounds or negative child interval, or negative parent duration, records a timestamp anomaly. Known missing parents and timestamp anomalies make fit evidence insufficient even if quantiles appear close. A capture may omit children without leaving a broken parent reference: completeness remains unknown and their unobserved time stays in the own-time estimate. Treat these estimates as conditional on the capture, not as measurements of CPU time.

Call inference still chooses sequential or parallel style by observed votes. Sequential generation sums child durations; parallel generation starts all children together and takes their maximum. Staggered overlaps, mixed styles, correlated calls, retries, async work, and varying repetition counts may not survive this approximation. Total-latency assessment exposes discrepancies; no new distribution or call-scheduling model is introduced.

Fit policy

Each own-time and total-time assessment records observation count, modeled count, observed and modeled median/p95 in nanoseconds, status, and reason codes. Quantiles use the nearest-rank definition. Statuses are:

• fit: both quantiles meet tolerance, with adequate evidence.
• poor_fit: median or p95 differs by more than tolerance.
• insufficient_evidence: fewer than 100 observed or modeled samples, known missing parents, timestamp anomalies, or bounded regeneration.

The per-quantile tolerance is the greater of 20% of the observed value or 1 ms. These are initial engineering thresholds, not statistical confidence intervals. At 100 observations, only about five observations describe the upper 5% tail; fewer samples are explicitly insufficient. The absolute floor avoids false precision for very short spans. For sub-millisecond systems, fit under this policy may be too permissive for your alert requirements.

Own-time assessment samples the actual parsed duration string after rounding and uses the existing normal sampler with zero clamping. Total-time assessment runs synth.GenerateTraces on the emitted execution model with seed 253 and up to 2,048 traces, collecting elapsed spans through its observer API, with a 100,000-span budget across the assessment. Budget exhaustion produces insufficient_evidence with an assessment_span_budget reason. Metadata records the span budget and actual modeled trace and span counts. Rarely visited operations may have insufficient modeled samples even in a large capture. The same fixed seed is used for 2,048 own-time draws per operation. The capture records these settings. No fit is promised to generalize to future traffic.

Validation cases establish the policy's practical meaning:

Case	Expected assessment
One 30 ms parent with a 10 ms child	Runnable 20 ms own-time estimate; insufficient evidence
100 such invocations	Own time and regenerated 30 ms totals fit
Sequential or aligned parallel children	Subtract union; restore original total
Children at 1–17 ms and 14–29 ms inside a 30 ms parent	Own-time fit; regenerated 18 ms total is a poor fit
Equal populations at 1 ms and 99 ms	Existing normal model is a poor fit
Positive duration below 1 microsecond	Emitted floor is assessed, rather than an unrounded estimate
Missing parent or out-of-bounds interval	Runnable estimate with insufficient evidence

Attributes and omissions

An operation retains a span attribute only when the same supported scalar, including its type, occurs on every observed span of that operation. Supported types are string, signed 64-bit integer, float64, and boolean. Integers are decoded without passing through a float; integer-valued floats retain explicit YAML float typing. Static attributes consume no random draws.

Each encountered key records presence count, unsupported count, status, retained scalar type, and zero or more reasons: partial_presence, varying_value_or_type, unsupported_value, or reserved_engine_attribute. Arrays, objects, binary values, nulls, and unsupported source types are omitted. Engine identity attributes (synth.service, synth.operation, synth.scenarios) are regenerated by Motel and reported as omitted.

Resource service.name supplies service identity. Other resource keys are counted under capture-wide resource_omissions, once per source span carrying that resource key. Span attributes are never promoted to resource attributes. Broader resource preservation and varying attribute models are deferred.

Reports and defaults

--report path.md renders the same structured evidence saved in the topology; the stderr summary counts operations by latency assessment. The report includes attribute omissions and capture settings. Evidence records observed source counts and the root observation window; traffic uses that window where possible and otherwise defaults to 1/s. Existing integer traffic-rate formatting remains an approximation. Inferred durations, calls, and probabilities remain editable execution fields. Evidence is a record of the original import only.

Operation import.inference also retains error counts, per-target call presence and occurrence counts, and parallel/sequential votes with the inferred style and its default-or-observed basis. requested_min_samples records the normalized --min-traces threshold. Structured reasons preserve low-sample and weak/mixed call-style warnings; terminal confidence diagnostics use this same assessment. The latency observation count supplies the parent count for call probabilities.