Retry overloaded responses, honour Retry-After, and log API errors

[olosegres]

Summary

• Retry 429 and 529 responses up to twice, waiting as long as Retry-After asks (exponential backoff when it is absent). Before, only 429 was retried, after a fixed 1s/2s pause, and a 529 went straight back to opencode. A requested wait longer than 10s is not absorbed inside the request: the response is returned at once so opencode's own retry loop — which shows the wait and can be cancelled — takes over, instead of a usage-limit 429 that resets in hours being retried twice for nothing. The wait ends early when the request is aborted, and 429 keeps its token refresh.
• Retry the OAuth token endpoint on 529 as well, honouring retry_after from its body, and return the endpoint's real last answer instead of a made-up "Rate limited" 429.
• Accept a plain-string system prompt. The request rewrite assumed text blocks; a string threw before the billing header was added, the error was swallowed, and the request went out unrewritten.
• Log failed responses. opencode only shows a one-line summary of a failed request; each failed response handed back to it (after retries) is now written to stderr and appended, with its request id, to ~/.local/share/opencode/log/claude-bridge-api-errors.log ($XDG_DATA_HOME is respected).

Test plan

• npm test — retry policy units; the plugin's fetch wrapper driven against a stubbed network (529 recovery, retries running out, long Retry-After, abort during the wait, string system prompt, error log); the token endpoint retry loop driven through a stand-in curl
• npm run typecheck
• With the plugin loaded in OpenCode, prompts go through as before, and a request for a nonexistent model leaves the expected 404 entry, with its request id, in the error log.