fix(stream): distinguish an API stream drop from a clean end / user Esc (#133)

[justrach]

Closes #133.

Problem

kimi-k2.6 (Moonshot, an OpenAI-compatible chat/completions provider) drops its SSE stream mid-generation — often mid-reasoning, before any output. The reporter saw "× interrupted (esc)" on 0.0.187, which predates the #134 fixes. Those fixes stopped an idle stall from being mislabeled as Esc, but left a residual gap for .openai-kind providers:

Once body bytes arrived (got_body), any later read close/reset was treated as a clean end (break :stream). isStreamEnd recognizes only the literal data: [DONE] sentinel for .openai, so a drop before [DONE] was silently accepted — and assembleOpenAI defaults the missing finish_reason to "stop", so a truncated answer was saved as a complete, successful turn with no signal at all (worse than the original symptom).

Fix — distinguish "confirmed complete" from "connection died"

• Add saw_done, set only when the provider's terminal event is actually observed: [DONE]/structural terminator (via isStreamEnd), or — for .openai — a non-null finish_reason chunk (openaiComplete), recorded without breaking early so a trailing usage chunk + [DONE] still land (no cost/usage regression).
• The three post-body read-close sites now break cleanly only if saw_done; a close with body but no terminal returns the new error.StreamDropped and prints ⚠ connection dropped — response ended early.
• error.StreamDropped is handled like a stall (no retry — the partial already streamed): agent_request records it + telemetry; mainloop/repl_glue keep the partial with [response ended early: connection dropped] (never [response interrupted by user]) and emit the structured --json error.

Safety

Genuine Esc handling is untouched (escPressed / esc_cancel / the stall watchdog). Documented behavior change: an OpenAI-compatible gateway that ends a stream with neither [DONE] nor a finish_reason (nonstandard) is now reported as a drop instead of silently succeeding — the partial is preserved either way.

Verified

zig build · zig build test 151/151 (new isStreamEnd/openaiComplete cases lock the drop-vs-clean-end distinction, incl. the kimi mid-reasoning case and an escaped-finish_reason-in-content guard) · a live codegraff/deepseek (.openai) turn still completes cleanly with no false "connection dropped".

Note: the original "interrupted (esc)" symptom is already fixed on main by the #134 commits; this closes the remaining silent-truncation gap for Moonshot-style mid-stream drops.

🤖 Generated with Claude Code