Add provider error taxonomy, model capabilities gating, and telemetry fixes

[TonsOfFun]

Summary

This PR implements three major framework improvements extracted from the ActiveAgents platform: a portable provider error taxonomy for cross-vendor exception handling, per-model capability gating to strip unsupported parameters before API calls, and fixes to telemetry instrumentation and redaction.

Key Changes

1. Provider Error Taxonomy (lib/active_agent/providers/errors.rb)

• Introduces ActiveAgent::Providers::Errors module with normalized exception classes: RateLimited, ContextLengthExceeded, AuthenticationFailed, ContentFiltered, ServiceUnavailable, and InvalidRequest
• Implements Taxonomy.normalize() that classifies vendor SDK exceptions by:
  • Demodulized class name (e.g., RateLimitError → RateLimited)
  • HTTP status code mapping (429 → RateLimited, 401/403 → AuthenticationFailed, etc.)
  • Message pattern matching for context overflow and content filter errors
• Preserves original exception as #cause for debugging
• Enables portable rescue_from policies across different LLM providers
• Integrated into with_exception_handling in the exception handler concern

2. Model Capabilities Gating (lib/active_agent/model_capabilities.rb)

• Adds ActiveAgent::ModelCapabilities registry to strip unsupported parameters before API calls
• Built-in rules for thinking-first Claude models (Opus 4.7+, Opus 5, Sonnet 5) and OpenAI reasoning models (o-series, GPT-5) that reject sampling parameters
• Extensible via ModelCapabilities.register(pattern, unsupported:) for custom/self-hosted models
• sanitize! method removes rejected parameters in-place and reports what was removed
• Can be disabled globally with ModelCapabilities.enabled = false
• Integrated into prepare_prompt_parameters in ActiveAgent::Base

3. Telemetry Improvements

• Redaction: Implements redact_spans in Tracer#build_trace_payload to scrub span and span-event attributes matching configured patterns before transmission/storage
  • Uses case-insensitive substring matching on attribute keys
  • Controlled by configuration.redact_attributes (defaults to common sensitive keys)
  • Applied to both span attributes and nested event attributes
• Instrumentation fixes:
  • Removes dead around_generate hook registration that was never called
  • Fixes provider_name and model_name attribute recording (now always set, not conditional on respond_to?)
  • Fixes messages.count recording by reading from prompt_options[:messages] instead of nonexistent messages method
  • Adds llm.provider and llm.model attributes to LLM spans

4. Tool-Loop Safety

• Adds max_tool_turns (default 25) to bound tool-calling recursion in BaseProvider
• Tracks tool_turns counter and stops processing when limit is reached
• Emits tool_turns_exceeded.active_agent notification when cap is hit
• Allows per-agent/prompt override via max_tool_turns: parameter

5. Documentation

• Adds docs/framework/v2-extraction-roadmap.md documenting the three-layer architecture and what belongs in each layer
• Clarifies which platform features are being absorbed into the framework vs. staying in the platform

Testing

Comprehensive test coverage added:

• test/providers/errors_taxonomy_test.rb — classification by name, status, message patterns; vendor exception wrapping
• test/model_capabilities_test.rb — parameter stripping, custom rules, sampling support detection
• test/providers/base_provider_tool_turns_test.rb — recursion bounding and notification emission
• test/dashboard/telemetry_redaction_test.rb — attribute scrubbing in spans and events
• test/dashboard/telemetry_correlation_test.rb — provider/model attribute recording in instrumented generations

Notable Implementation

https://claude.ai/code/session_01B5PaDq3uaUWuA8UW4K7siP