feat: configurable retryPolicy and honour all_proxy in env proxy

[AlienHub]

Problem

Two pain points when using the openai-codex route on real networks:

1. Retry count is not configurable. CodexAuthAdapter hardcodes
   resolveRetryPolicy(undefined, …), so the route always uses the harness
   normal policy (max 2 retries). On a flaky/rate-limited
   chatgpt.com/backend-api, requests frequently need more attempts and there
   is no way to raise the budget without editing the installed package.

2. all_proxy is ignored. installEnvHttpProxy only reads
   HTTPS_PROXY/https_proxy/HTTP_PROXY/http_proxy. Many proxy setups
   (e.g. Clash) set only all_proxy, which curl and the codex CLI honour —
   but Node's undici does not, and new EnvHttpProxyAgent() also reads only
   http_proxy/HTTPS_PROXY. On such networks every request to
   chatgpt.com went direct and timed out (connect timeout), while the
   browser/CLI worked fine.

Changes

• Configurable retry policy: optional retryPolicy on the
  llm-codex-auth row (validated with RetryPolicySchema), threaded through
  CodexAuthAdapterOptions into resolveRetryPolicy. Omitted ⇒ unchanged
  default (normal, 2 retries). Example:

  - id: llm-codex-auth
    config:
      retryPolicy:
        mode: normal
        maxRetries: 5

• all_proxy support: the env-proxy chain now includes
  ALL_PROXY/all_proxy, and the resolved URL is passed explicitly as
  { httpProxy, httpsProxy } (undici's env parser never reads all_proxy).
  NO_PROXY behaviour is preserved from the environment.

• README (en/zh) config table and CHANGELOG updated.

Validation

• pnpm run check passes: oxlint, typecheck (both host and client, with
  exactOptionalPropertyTypes), vitest, build, package smoke, publint.
• On a machine where all_proxy=http://127.0.0.1:7890 is the only proxy env
  var: undici direct to chatgpt.com/backend-api times out; with the patched
  dispatcher the same request succeeds through the proxy (HTTP 401, as
  expected without auth headers).

[suntianc]

Thanks for the contribution — both directions are useful. Configurable retry behavior is valuable on flaky/rate-limited networks, and supporting ALL_PROXY/all_proxy addresses a real gap between Codex CLI/curl and Undici.

A few changes are needed before this can merge:

1. Please rebase onto the current main. The PR currently conflicts in CHANGELOG.md and src/codex-auth-adapter.ts. When resolving the adapter conflict, please preserve the newer auth: codexAuthInjection() and maxRequestImageBytes behavior from main.
2. Preserve protocol-specific proxy semantics. The current code resolves one URL and passes it as both httpProxy and httpsProxy. This collapses distinct http_proxy/https_proxy settings, can route HTTP through the HTTPS proxy, and changes Undici's lowercase-before-uppercase precedence. Please resolve HTTP and HTTPS independently, using all_proxy/ALL_PROXY only as the fallback for each, while continuing to leave NO_PROXY to EnvHttpProxyAgent.
3. Update the documented default. Current main uses @deepseek-ai/dsh-llm@0.1.1-rc.1, where resolveRetryPolicy(undefined) defaults to normal mode with maxRetries: 5, not 2. Delegating the omitted value is still correct; the README, changelog, and source comments should say 5 or avoid hard-coding the number.
4. Please add tests for retry-policy schema/default/forwarding and for proxy precedence, all_proxy fallback, and NO_PROXY preservation. The isolated PR-head pnpm run check passes, but the new behavior is currently untested.

So: the feature direction is worth keeping; the main blockers are the stale base and proxy-resolution regression.