Add transport middleware pipeline with retry, rate limit, and metrics

[briansrls]

Summary

This PR introduces a comprehensive middleware infrastructure for the transport layer, enabling cross-cutting concerns like automatic retry with exponential backoff, rate limiting with token bucket/sliding window algorithms, circuit breaker protection, and observability through metrics collection.

Key Changes

• Middleware Infrastructure (middleware/mod.rs): Core trait-based middleware system with TransportMiddleware trait, MiddlewareContext for request metadata, and PostProcessOutcome for handling retries/aborts.

• Retry Middleware (retry.rs):

  • Automatic retry for transient failures (5xx, network errors, rate limits)
  • Configurable backoff strategies: Fixed, Exponential, and ExponentialJitter
  • Circuit breaker state machine (Closed → Open → HalfOpen) to prevent cascading failures
  • Safety checks: only retries idempotent or readonly operations by default

• Rate Limit Middleware (rate_limit.rs):

  • Token bucket and sliding window rate limiting algorithms
  • Per-scope rate limit tracking (e.g., "github:core", "github:search")
  • Respects Retry-After headers from 429 responses
  • Headroom tracking for metrics

• Metrics Middleware (metrics.rs):

  • Pluggable MetricsSink trait for metrics collection
  • Built-in implementations: NullMetricsSink, LogMetricsSink, InMemoryMetricsSink
  • Records request/response timing, retry attempts, rate limit headroom, and error classifications

• Pipeline Composition (pipeline.rs):

  • TransportPipelineBuilder for composing middleware layers
  • TransportPipeline for executing requests through the middleware stack
  • Standard pipeline builder that automatically configures metrics, rate limit, and retry based on config
  • Proper request/response flow: outer→inner on request, inner→outer on response

• Classification Integration (classify.rs):

  • New classify_for_middleware() function for middleware to classify responses
  • Helper functions: extract_status_code(), is_success()
  • Supports both HTTP and non-HTTP transport types

Notable Implementation Details

• Safety-First Retry: Non-idempotent operations require explicit configuration to retry, preventing accidental duplicate operations
• Circuit Breaker: Protects against cascading failures with configurable thresholds and recovery probes
• Flexible Rate Limiting: Supports multiple algorithms and respects provider-specific rate limit headers
• Composable Metrics: Metrics sink is pluggable, allowing integration with various observability systems
• Deterministic Jitter: Exponential jitter uses deterministic calculation for reproducibility in tests
• Proper Middleware Ordering: Metrics (outermost) → rate limit → retry → execute, ensuring correct timing and classification

All middleware is fully tested with comprehensive unit tests covering success paths, error conditions, and edge cases.

https://claude.ai/code/session_01DwRsV4WNLUPLdv6Y4ucaGS