Retry strategies

Strategies can use one of two signatures:

• Context-aware: strategy(ctx: BackoffContext) -> float
• Legacy: strategy(attempt: int, klass: ErrorClass, prev_sleep_s: float | None) -> float

BackoffContext provides:

• attempt – 1-based attempt number
• classification – structured Classification (retry hints, details)
• remaining_s – deadline time remaining (seconds)
• prev_sleep_s – previous sleep value
• cause – "exception" or "result", indicating whether the failure came from a raised exception or a classified return value

Built-ins (all return a StrategyFn):

• decorrelated_jitter(base_s=0.25, max_s=30.0) – sleeps uniformly in [base_s, prev_sleep*3], clamped to max_s.
• equal_jitter(base_s=0.25, max_s=30.0) – exponential with jitter in [cap/2, cap], cap = min(max_s, base_s*2^attempt).
• token_backoff(base_s=0.25, max_s=20.0) – gentler exponential (1.5^attempt) with jitter.
• retry_after_or(fallback, jitter_s=0.25) – honors Classification.retry_after_s when present, otherwise calls fallback.
• adaptive(fallback, window_s=60.0, target_success=0.9, min_multiplier=1.0, max_multiplier=5.0) – wraps a fallback strategy with adaptive scaling based on recent success/failure rate (thread-safe).

Example:

from redress import ErrorClass, RetryPolicy, default_classifier
from redress.strategies import decorrelated_jitter, equal_jitter

policy = RetryPolicy(
    classifier=default_classifier,
    strategy=decorrelated_jitter(max_s=10.0),  # default for most classes
    strategies={
        ErrorClass.RATE_LIMIT: decorrelated_jitter(max_s=60.0),
        ErrorClass.SERVER_ERROR: equal_jitter(max_s=30.0),
    },
)

Retry-After-aware example:

from redress import RetryPolicy
from redress.extras import http_retry_after_classifier
from redress.strategies import decorrelated_jitter, retry_after_or

policy = RetryPolicy(
    classifier=http_retry_after_classifier,
    strategy=retry_after_or(decorrelated_jitter(max_s=10.0)),
)