fix(agent): don't peek past a completed stream on a half-open socket (#56)

[justrach]

Problem

The SSE streaming reader's post-line probe, reader.peekByte() (agent_stream.zig), ran before the terminal-event check that breaks on [DONE] / response.completed. On the final line of a response the peek's result is discarded anyway (the loop breaks at the terminal check), but its fill can block forever on a half-open socket — server gone, no FIN/RST — which the idle-stall watchdog does not guard. That hangs an already-complete turn: part of #56's "reconnect doesn't reconnect".

Fix

Move the peek to after the terminal-event break, which matches this loop's own documented intent — "stop instead of waiting for the socket to close". A completed response now breaks before it ever peeks; the non-terminal path is byte-for-byte unchanged (peek → if (!more) break → toss the buffered '\n').

Safe by inspection: on a terminal line the loop already broke at the isStreamEnd check, so more was never read there — moving the peek only removes a useless, hang-prone fill on completed responses.

Verification

• zig build test green.
• Streaming PTY suite green — test-pty-codex-ws, test-pty-repl, test-pty-markdown, test-pty-spinner (these stream model output through this exact reader), so no regression in the hot read path.

I explored a mock hang-test but it can't reliably distinguish old-from-new: a complete SSE line always leaves its terminating \n buffered (that's the byte reader.toss(1) consumes), so the old peekByte reads from the buffer without blocking in every reproducible case. A test both versions pass isn't a meaningful guard, so I rely on the inspection argument + regression suite instead.

Scope (honest)

• This PR closes the post-completion subcase (the documented "don't wait on the socket" intent).
• The mid-stream half-open subcase — peekByte blocking on a non-terminal line — still needs a watchdog-guarded peek (a streamPeekTask select-arm) plus a half-open-socket mock harness (accept, send partial SSE, then network-drop without closing) to verify. Left as a follow-up.
• Fix-B from the issue (retryable stalls via watchdogError: Esc→Interrupted, deadline→HungRequest/StreamStalled) is already implemented and tested (Stream stall incorrectly reported as user Esc interruption #134).

Pre-existing zig fmt debt in this file (unrelated one-liners) left untouched to avoid churn.

Refs #56