Fix Langfuse v4 observation semantics

[ChoMinGi]

What changed

Align the 0.2 line with Langfuse v4 observation semantics. Root and generic spans now remain spans, model operations emit their explicit observation types, root I/O uses observation attributes, and trace-wide fields propagate safely across nested and cross-thread work.

The Spring AI and LangChain4j adapters now emit explicit model types. This also adds an opt-in exact-SHA release canary, fixes the tagged live-test selector so zero executed tests fail, and documents the actual proxy, async, and external-OpenTelemetry boundaries.

Live Langfuse export and API read-back are intentionally still a pre-tag release gate. They require the final commit SHA and release credentials, so this PR does not claim that gate has run.

Testing

• ./mvnw -B -ntp clean verify — core 119 tests; starter 193 tests with 2 conditional skips; packaging, coverage, and API compatibility passed
• quality profile — dependency convergence, license checks, SBOM generation, and SpotBugs passed
• Spring AI consumers passed with 1.0.9 and 1.1.8
• LangChain4j consumers passed with 1.0.0 and 1.18.0
• both examples and the core prompt consumer passed against the locally installed artifacts
• actionlint, bash -n, shellcheck, and git diff --check passed
• release-canary selection was verified to execute exactly one test and fail without credentials rather than succeeding with zero tests

Checklist

• ./mvnw -B -ntp clean verify passes
• Applicable independent builds under consumer-tests pass after a local reactor install
• Tests cover the changed behavior
• Public API or user-facing behavior is documented
• No credentials or captured model data are included