fix: align API retry logic with Claude Code (Retry-After header, jitter, 10 retries)

[lewis617]

Summary

Align API retry logic in openaiClient.ts with Claude Code's withRetry.ts to avoid ineffective retries during rate limiting.

Changes

• Honor Retry-After response header: when the server returns 429 with a Retry-After header, wait the specified duration instead of using exponential backoff. This was the root cause of the first 3 retries being wasted — the server hadn't lifted the rate limit yet.
• Increase max retries: 5 → 10
• Faster initial retry with jitter: base delay 2000ms → 500ms, with 25% random jitter to avoid thundering herd
• Max delay cap: 32s (prevents unbounded waits)
• Log improvement: include retryAfter value in warning logs

Spec

Updated specs/012-ai-error-handling.md:

• User story 3 expanded to 3 acceptance scenarios (Retry-After priority, exponential backoff+jitter, 5xx shared strategy)
• FR-005 rewritten with full retry parameters
• FR-010 simplified to reference FR-005

Test

• Updated exponential backoff test for new delay sequence (500→1000→2000→...→32000)
• Updated max retries test (5→10, 6→11 total calls)
• Added Retry-After header test: verifies 30s header is honored and no retry happens before that
• All 15 tests pass, type-check and lint pass

Before vs After

	Before	After
Retry-After header	Ignored	Honored (bypasses backoff)
Max retries	5	10
Base delay	2000ms	500ms
Jitter	None	25%
Max delay	None	32s