feat(loops+otel): nested GenAI-semconv span tree for loop topology (Tangle Intelligence viewer)

[tangletools]

Makes the driven/dynamic loops render as a real topology tree in any
OpenTelemetry/GenAI viewer (Tangle Intelligence in ADC, Phoenix, Tempo, …) over
the existing OTLP export path — no new channel.

Problem

Loop traces were flat, zero-duration point spans with bespoke loop.*
attributes: ingestible, but renders as a flat list, and the agent-authored
topology (the per-round move + why) was invisible — the driver/planner never
reach the trace layer.

Future-proofing — standard, not bespoke

Spans now follow the current OpenTelemetry GenAI semantic conventions
(registry),
explicitly avoiding the deprecated keys:

• ✅ gen_ai.operation.name (invoke_workflow / invoke_agent), gen_ai.agent.name,
  gen_ai.conversation.id, gen_ai.usage.input_tokens / output_tokens
• 🚫 not gen_ai.system, gen_ai.usage.prompt_tokens, gen_ai.usage.completion_tokens
• tangle.loop.* / tangle.cost.usd namespaced extension for what OTel hasn't
  standardized (topology move, verdict, placement, cost) — the OTel-blessed
  vendor-extension escape hatch, so a future standard is a one-namespace remap.

Shape

loop (invoke_workflow)
  └─ loop.round[k] (invoke_workflow)   tangle.loop.move.{kind,width,rationale}, .decision
       ├─ loop.iteration[i] (invoke_agent)  gen_ai.agent.name, gen_ai.usage.*,
       │     tangle.loop.verdict.{valid,score}, tangle.loop.placement.kind, tangle.cost.usd
       └─ …

Changes

• Kernel: new loop.plan event per round ({roundIndex, plannedCount, moveKind, rationale})
  via an optional Driver.describePlan() (dynamic driver returns its move's kind+rationale;
  refine/fanout infer from count). loop.iteration.ended now carries tokenUsage.
• otel-export: buildLoopOtelSpans(events, traceId, parentSpanId?) — pure builder of the
  nested, real-duration tree with the attribute schema above. loopEventToOtelSpan (flat)
  kept for back-compat. Scope version 0.23.0 → 0.33.0.
• trace-propagation: buffers events per runId, flushes the tree on loop.ended so the
  live MCP→OTLP path ships the hierarchy.

Tests

buildLoopOtelSpans tree shape + real durations + gen_ai/tangle attrs + a
no-deprecated-keys assertion (+4); kernel emits loop.plan with move kind +
rationale and iteration tokenUsage (+1). Full suite 403 green, tsc + biome clean.

Increment 2 (separate): the planner's own gen_ai span (needs
PlannerContext.traceEmitter). Serves #754 (intelligence ingestor).