Unify retry systems into a single RetryPolicy

[thedumbtechguy]

Summary

ChronoForge had three independent retry systems, two backoff algorithms, and three "should we retry?" decision models:

1. Workflow-level (uncaught errors in perform) — should_retry? hardcoded to attempt < 3, backoff from a fixed [1s,5s,30s,2m,10m] array (whose 2m/10m tail was unreachable because the < 3 check bit first), guarded by a contradictory max_attempts == 5.
2. Step-level (durably_execute, durably_repeat) — max_attempts: param with a different algorithm (2**n capped at 32s).
3. wait_until — retry_on: [classes] allowlist, no count cap.

This PR collapses all of them into one RetryPolicy (attempt cap + exponential-with-jitter backoff + error-class predicate). Today's behaviors become presets of one type.

What changed

• New ChronoForge::Executor::RetryPolicy — retryable?(error, attempts) and backoff_for(attempts) are the entire decision surface. Backoff is computed once at re-enqueue (never replayed, so deterministic where it matters).
• Per-call retry_policy: on durably_execute, durably_repeat, wait_until, plus a class-level retry_policy DSL. Resolution: per-call → class default → per-site built-in.
• Per-site built-ins: steps 3 attempts / cap 30s; workflow-level 10 attempts / cap 600s (~8.5 min tolerant window for transient infra errors on uncaught perform errors); wait_until retries nothing by default. wait_until deliberately does not inherit the class default, so a class-wide "retry everything" can't silently retry condition-evaluation bugs.
• Deletions: RetryStrategy, the dead should_retry? (the 3-vs-5 contradiction), and the dead retry_method: reschedule arg.

Notes / decisions

• The kwarg is retry_policy: (not retry:) because retry is a Ruby keyword — a retry: param can't be read inside a method without binding.local_variable_get(:retry).
• Workflow-level retries fire only on uncaught errors in perform; a step exhausting its own retries stalls the workflow instead (unchanged). The 10 was chosen to ride out realistic transient blips (DB failover, deploy restart) without dragging out the deterministic-bug case — each workflow-level retry replays the whole workflow.

⚠️ Breaking changes

• durably_execute/durably_repeat no longer accept max_attempts:; wait_until no longer accepts retry_on:. All three take retry_policy: now. Migrate max_attempts: N → retry_policy: RetryPolicy.new(max_attempts: N) and retry_on: [...] → retry_policy: RetryPolicy.new(retry_on: [...]).
• Backoff is exponential-with-jitter everywhere. A permanently-failing workflow now reaches failed after 10 attempts (was an effective 4).

Testing

• New RetryPolicy unit tests (truth table for retryable?, backoff growth/cap/jitter bounds, presets) and integration tests (per-call override, class default, wait_until fast-fail vs opt-in retry).
• Full suite green: 105 tests, 0 failures; standardrb clean.

A design doc is included at docs/superpowers/specs/2026-06-03-unified-retry-policy-design.md.

🤖 Generated with Claude Code