fix: retain TLS output and causes after send failure

[kentbull]

What

• Retry typed TLS WANT conditions without changing terminal state.
• Record broken pipe as transmit-only closure so receive can still drain.
• Force-close both directions for fatal SSL and connection-wide socket failures.
• Retain the original exception and every unsent byte on terminal paths.

Why

A nonblocking retry is not a failure, a broken write does not necessarily end receive, and an unusable TLS transport must not remain represented as open. Conflating those outcomes either loses accepted output or permits impossible retries.

Boundary

This PR classifies established-session TLS send outcomes only. It does not initiate close_notify or implement recurrent graceful shutdown.

[kentbull]

Upstream prerequisites: the directional state and unsent-output contract from #9/#10, followed by the truthful TLS wrapping and receive-outcome classification from #11.

Only then can this send-side behavior classify WANT, broken pipe, fatal TLS, and connection-wide socket failures while retaining the original cause and unsent bytes.

[kentbull]

Implements downstream send-side behavior for ioflo#174 by separating retryable WANT conditions, broken pipe, fatal TLS errors, and connection-wide socket failures while retaining terminal causes and unsent plaintext. #16 expands regression coverage.