fix(feishu): add transient network retry for message send operations

[0xsegfaulted]

Summary

• Add withTransientRetry wrapper with exponential backoff + jitter (500ms→1s→2s, max 3 retries, +25% jitter) for all Feishu HTTP API calls
• Detect transient network errors via typed syscall checks (ECONNRESET, EPIPE), net.Error timeout, io.EOF, and string-matching fallback
• Wrap all 8 Feishu API call sites: reply, create, patch, delete, upload image/file/audio, and preview start

Problem

When Feishu WebSocket connections experience rapid resets (every 20-40s for ~3 minutes), HTTP API calls for sending messages also fail with transient network errors (EOF, connection reset by peer, i/o timeout). Previously these errors were not retried — only token refresh retry (code 99991663) existed — causing messages to be silently lost with delays of 5-25 minutes.

Design

• Layering: transient retry (outer) wraps token refresh retry (inner), so a transient retry can re-trigger a fresh token acquisition if needed
• Jitter: +25% random jitter on each backoff delay to prevent thundering-herd retries from concurrent sessions
• Non-transient errors (API error codes, rate limits, validation) are not retried — returned immediately

Test plan

• TestIsTransientError — 16 error type subtests including syscall, net.Error, EOF, string matching, and negative cases
• TestWithTransientRetry_* — 5 unit tests: success, retry-then-succeed, non-transient passthrough, max retry exhaustion, context cancellation
• TestReplyRetriesOnTransientNetworkError — integration test with httptest + connection hijacking
• TestCreateMessageRetriesOnTransientNetworkError — integration test
• TestPatchMessageRetriesOnTransientError — integration test
• TestReplyDoesNotRetryOnNonTransientAPIError — integration test
• TestReplyTransientRetryThenTokenRefresh — validates both retry layers working together
• TestWithTransientRetry_BackoffTiming — verifies backoff timing is reasonable
• All existing feishu tests still pass
• go build ./... clean

🤖 Generated with Claude Code

[chenhg5]

LGTM. Well-structured transient network retry implementation for Feishu.

Review summary:

• ✅ withTransientRetry wrapper with exponential backoff + jitter (500ms→1s→2s, max 3 retries, +25% jitter)
• ✅ Comprehensive transient error detection: syscall checks (ECONNRESET, EPIPE), net.Error timeout, io.EOF, string-matching fallback
• ✅ Layered design: transient retry (outer) wraps token refresh retry (inner)
• ✅ Non-transient errors (API errors, rate limits) are not retried
• ✅ All 8 Feishu API call sites wrapped
• ✅ Comprehensive test coverage: 16 error type subtests, integration tests with httptest + connection hijacking
• ✅ CI passes (lint, unit-test, smoke-test, regression-test, performance-test)

Great contribution! Thanks for the detailed PR description and comprehensive tests.