feat(observability): expose context and cache metrics

[wuyuanfr]

Summary

• emit content-free final-context evidence, payload fingerprints, budget and compression metrics through OpenTelemetry
• forward compression and provider prompt-cache metrics through the observer, including DeepSeek cache-token reporting
• emit direct tool return values to execution logs and preserve ToolConfig input/output schemas
• bound ContextManager evidence fingerprint recursion with cycle, depth, and normalization-error fallbacks

Context architecture adaptation

The original observability changes predated the unified ContextManager runtime introduced by #3475. This PR ports the metrics onto the current ContextEvidence and ContextEvidenceCollector path instead of restoring the removed legacy runtime. Existing soft/hard budget, raw/final token, processing-mode, and compaction fields remain the source of truth.

Mixed change disclosure

The historical 6b01d9310 change is intentionally included in full. In addition to observer instrumentation, it restores ToolConfig.inputs and output_type on instantiated tools when decorated callables are not introspectable by smolagents.

Tests

• 109 context, prompt-cache, and monitoring tests passed
• 4 targeted NexentAgent tests passed
• 3 targeted CoreAgent tests passed

缓存命中率指标观测：

压缩相关指标观测：

具体trace观测内容：

[codecov]

Codecov Report

❌ Patch coverage is 92.20779% with 6 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
sdk/nexent/core/agents/nexent_agent.py	76.00%	2 Missing and 4 partials ⚠️

📢 Thoughts on this report? Let us know!