fix(weixin): detect stale context_token when iLink returns "rate limited" (#35949)

[PRATHAMESH75]

Summary

Fixes #35949 — Weixin iLink stale context_token was misidentified as a rate limit, causing permanent outbound send failure.

iLink's sendmessage endpoint returns ret=-2, errmsg="rate limited" for an expired context_token, not just "unknown error". The stale-session matcher
_is_stale_session_ret only recognised "unknown error", so every outbound message exhausted all 4 retries against the dead token. Only inbound replies survived because they
refresh the token from the incoming webhook.

Fix

case-insensitive comparison preserved).

The send loop at _send_text_chunk (weixin.py:1640) already guards itself with retried_without_token, so the false-positive cost on a genuine rate-limit is capped at one
tokenless retry before falling through to the rate-limit backoff branch. Inbound paths and the -14 SESSION_EXPIRED_ERRCODE path are untouched.

Chose Option B from the issue (expand the matcher) over Option A (treat all ret=-2 as stale) to preserve the existing backoff behavior for genuine throttling.

Test plan

• New regression test TestIsStaleSessionRet::test_ret_minus_2_with_rate_limited_is_stale covers "rate limited", mixed case, and surrounding whitespace.
• Existing 7 tests in TestIsStaleSessionRet still pass (no regression on "unknown error", "freq limit", empty errmsg, -14, or success codes).
• pytest tests/gateway/test_weixin.py::TestIsStaleSessionRet — 8 passed.
• Manual: confirm stale-token chat now recovers on next send instead of looping with ret=-2.

Workaround (still valid for users on older builds)

Delete the cached token file to force regeneration.

[teknium1]

Thanks for tracing the stale-token path; current main still has the reported outbound classification gap.

Problems

• gateway/platforms/weixin.py:99 is shared by both outbound sends and long polling. The proposed "rate limited" match at PR line 113 also feeds getUpdates at gateway/platforms/weixin.py:1360, which treats a match as session expiry and sleeps for 600 seconds at lines 1361-1364. The send-only retried_without_token guard does not protect that path.

Suggested changes

• Scope the new "rate limited" interpretation to _send_text_chunk_locked (gateway/platforms/weixin.py:1772-1788) so polling retains its existing retry/backoff behavior.
• Add a send-path regression asserting a "rate limited" response retries once with context_token=None, plus a polling-path regression for the same response.

This is an automated hermes-sweeper review.

[teknium1]

_is_stale_session_ret is also called by the getUpdates loop. This makes any long-poll ret=-2, errmsg="rate limited" response take the session-expired branch and sleep 600 seconds; the tokenless retry guard cited here exists only in the send path. Please scope this extra phrase to outbound sendmessage handling.