[khunj]

What

Implements the A2A-T Java SDK observability layer (Plan A of the Java port of the a2a-t-sdk-python observability module), per the design doc docs/superpowers/specs/2026-10-20-observability-java-port-design.md:

• New module �2a-t-observability: A2ATObservabilityConfig (explicit > .env > env > defaults, 14 env keys), ObservabilitySetup (idempotent OTel 3-signal auto-setup: OTLP gRPC/HTTP + in-house Console fallback), manual span API (A2ATSpan/CurrentSpans), W3C raceparent propagation (TracePropagation), payload channels (PayloadExtractor + PayloadSanitizer, redact-then-truncate), LLM token metric + ThreadLocal usage stash (ObservabilityLLMClient), negotiation rounds metric (NegotiationMetrics), structured JSON logs on SLF4J + OTel Logger dual channel (ObservabilityLogs), and attribute constants/extractors byte-identical to the Python SDK naming registry.
• New module skeleton �2a-t-observability-a2a: hosts the future a2a-java protocol-level decorators (Plan B, separate PR); pins a2a-java 1.1.0.Final in its own pom so root dependencyManagement cannot break sample transitive resolution.
• �2a-t-core: TelemetryHook/TelemetryOperation SPI with a NO_OP default — telemetry can never break the business flow (all callbacks guarded, WARN + fallback).
• �2a-t-client / �2a-t-server: secondary constructors A2ATClient(Path, @nullable TelemetryHook) / A2ATServer(Path, @nullable TelemetryHook); all 45 public facade methods wrapped as �2at.sdk.{role}.{method} spans; terminal negotiation rounds reported from continueNegotiation (AGREED→accept, REJECTED→reject).
• env.example: observability section (14 keys).

Why

The Python SDK ships a full observability layer (spans, metrics, logs, propagation) targeting a2a-java dashboard parity; the Java SDK had none. This port keeps every span name, attribute key, metric name/unit and log event byte-identical to the Python SDK so existing dashboards work unchanged, while adapting the architecture to Java idioms: OTel api as a compile dependency with sdk/exporters , a hook SPI instead of instance duck-typing, and ThreadLocal instead of contextvar for the LLM usage stash.

Notes

• OTel 1.64.0 API adaptations (javap-verified): GlobalOpenTelemetry.getOrNoop() everywhere except the single intentional set() in setup (prevents get() auto-set poisoning that would silently disable auto-setup); no Java-21 pattern switches;
  esetForTest(); in-house Console exporters (opentelemetry-exporter-logging is not vendored).
• Verified: 9-module regression green (�2a-t-sample has a pre-existing AgentCard compile error unrelated to this PR and predating this branch; �2a-t-corpus is test-asset only).
• Deferred follow-ups and Plan B inputs are recorded at the end of the plan doc (docs/superpowers/plans/2026-10-20-observability-core.md).

Type of Change

• Bug fix (non-breaking change that fixes an issue)
• New feature (non-breaking change that adds functionality)
• Breaking change (fix or feature that would cause existing functionality to change)
• Documentation update
• Other (corpus test-asset addition: input case JSONs only, no SDK source change)

Checklist

• I have read the CONTRIBUTING guide
• All commits are signed off (git commit -s) per the DCO
• New and existing tests pass locally
• I have added tests that prove my fix is effective or my feature works (if applicable)
• I have added or updated documentation as needed
• New source files include the SPDX license header (repo-wide convention carries no SPDX headers; pom declares Apache-2.0 — will add on reviewer request)