fix: retry verified mislabeled SSE disconnect

[etwk]

Change

• Treat an Anthropic SSE invalid_request_error as transient only when its complete message exactly matches the verified transport-disconnect signature from issue Buffered transport disconnect mislabeled invalid_request_error bypasses retries #1.
• Preserve permanent handling of invalid schemas, malformed tool messages, context-limit failures, and near-match messages.
• Use the existing retry count, deadline, and buffering limits; discard failed buffered output and never transparently replay after commit.
• Add classification and HTTP regression coverage, and document the exact-match compatibility exception.

This is an implementation-specific compatibility exception, not a general transport-error detector. It does not change retry budgets or timeouts.

Verification

• Focused regression tests passed.
• go test -race ./... -count=1 passed.
• go vet ./... passed.
• A compiled proxy passed nine synthetic HTTP smoke scenarios: buffered recovery in two attempts; exhausted and disabled local retries; insufficient deadline and buffering window; schema, tool-message, and context-limit errors; and no replay after client-observed content.

Only synthetic protocol data was used for verification; no production request content or deployment details are included.

Fixes #1