…_DISPLAY

test_clipboard_target_linux_cli temporarily mutates the process-wide PATH
and WAYLAND_DISPLAY env vars to mock wl-copy. Running concurrently with
test_clipboard_target_success let the latter observe the mock during
deliver() but not during the follow-up test() call once the other test
had restored PATH, causing a flaky "assertion failed: test_res.reachable"
failure in CI. Serialize both tests with a mutex.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PThnjsJtj5TRZNBzCLECoi

Lets a hotkey deliver transcribed text to any OpenAI-compatible chat
completions endpoint and route the model's response (e.g. a summary or
rewrite) to the target's destination, instead of just the raw transcript.
Reuses the global OpenAI API connection settings (endpoint/api_key/model)
with optional per-target overrides for model, max_tokens, and timeout.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PThnjsJtj5TRZNBzCLECoi

Config is wrapped (state.config.lock().await.data.*), not flat;
the callback registered for the new OpenAI API target referenced
.ollama directly instead of .data.ollama, breaking the build.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PThnjsJtj5TRZNBzCLECoi

cargo test failed to compile voxctrl-app because two OutputTarget
struct literals in the sequential-delivery test predated the new
openai_* fields.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PThnjsJtj5TRZNBzCLECoi

Each openai_api target now keeps an in-memory chat log (last 20
messages / 10 exchanges) and replays it with every request, so
follow-up dictations like "what did I just ask you?" resolve
correctly instead of being treated as an isolated one-shot prompt.
History lives only in the running process and resets on hot-reload
or restart.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PThnjsJtj5TRZNBzCLECoi

The earlier openai_api target history fix didn't help users whose Q&A
flow goes through the legacy ollama_enabled pre-processing pipeline
(OllamaClient::process), which was completely stateless. Add a
process_with_history method and a per-binding in-memory history map on
InferenceEngine so follow-up questions like "what did I just ask you"
resolve correctly for that pipeline too.