Reusable Python Retry and Backoff Decorator Library Prompt

You are a senior Python engineer who builds resilience primitives that are correct, observable, and impossible to misuse into a retry storm.

I will provide:
- The kinds of failures I want to retry (exception types, predicate on the return value)
- Sync vs async functions
- Constraints (max attempts, max total time, idempotency guarantees)

Build a `retry` decorator library (a focused module, not a dependency on a heavy framework) that:

1. **Works as a configurable decorator** — `@retry(attempts=5, on=(ConnectionError, TimeoutError), backoff=exp(base=0.5, cap=30), jitter=full)`. Sensible defaults; everything overridable.

2. **Supports exponential backoff with jitter** — implement full jitter (random between 0 and the computed cap) to avoid thundering herds. Provide constant/linear/exponential strategies as small composable callables.

3. **Filters precisely** — retry only on the configured exception types or a `retry_if(result)` predicate. Never blanket-`except Exception`. Re-raise non-matching errors immediately.

4. **Bounds everything** — max attempts AND optional max total elapsed time (deadline). Stop at whichever comes first; raise the last exception (or a `RetryError` wrapping it) with attempt count.

5. **Handles sync and async** — detect coroutine functions and provide an async path that `await asyncio.sleep`s for the backoff (never blocks the loop). Same API for both.

6. **Is observable** — a `before_sleep` hook (log attempt, exception, next delay) and a `on_giveup` hook. Default logging is informative but quiet.

7. **Respects Retry-After** — allow a hook to override the next delay from an exception attribute (e.g. an HTTP 429's Retry-After).

8. **Preserves the wrapped function** — use `functools.wraps`; keep signature/typing intact.

Output: (a) the decorator + backoff strategy callables, (b) sync and async paths, (c) the `RetryError`, (d) `pytest` tests asserting attempt counts, exception filtering, deadline behavior, and jitter bounds (with a patched sleep), (e) a usage example.

Bias toward: precise exception filtering, full jitter, hard deadlines, and a default that warns the user this should only wrap idempotent operations.