NurCLI v0.43.1

Perplexity works again, a cloned repository can no longer approve its own tool calls, and the first request of a turn leaves seconds sooner.

• Perplexity moves to the Agent API. Perplexity retired Sonar's Chat Completions endpoint on 2026-09-27, which broke the perplexity route. It now speaks Responses at https://api.perplexity.ai/v1 and starts on perplexity/sonar; the models Perplexity hosts (anthropic/..., openai/..., google/..., xai/..., perplexity/...) work by id. nur fits every request to the Agent API's stricter schema: no replayed reasoning from other vendors, plain-text assistant turns, image URLs within its 2048-character cap, effort from minimal to max, and an output ceiling for Anthropic models, which require one.
• Project permission rules need your trust. A repository's .nur/permissions.toml used to take effect the moment you opened it, so allow = ["bash"] in a cloned repo ran its commands without asking, even in manual mode. Its deny and ask rules still apply at once, but its allow rules now wait for /permissions trust (or nur permissions trust), and any later edit to that list holds it again. nur tells you at startup what a project asks to auto-approve (nur run prints it to stderr). /permissions untrust withdraws trust, and /cd now reloads the rules for the directory you move to.
• Faster first request. Before its first model request, a session's first turn launched npm prefix -g while looking for Python, and every turn launched the PLUR CLI. Binary lookup now asks PATH before probing runtimes, the TUI fetches PLUR's context and OptMem's wake while you type, and later turns reuse them while PLUR refreshes in the background. Measured with release builds on Windows: Enter to request in the TUI went from about 2.1 s to 0.27 s, and a warm headless nur run from about 2.3 s to 1.3 s.
• Demo pipeline out of the repository. The film's recorder, player and renders no longer ship in the repo; the README shows the poster hosted on nuroctane.xyz/cli, which plays the film.
• Fixed: NUR_STARTUP_TRACE rows written from two threads at once could interleave mid-line, which made the startup suite fail now and then. Each row is now written in one piece.
• E2E: three new scenarios cover the Perplexity wire shape and untrusted and trusted project rules. The suite's warm skill index never matched a new scenario's home, so every scenario rescanned every skill root; it now carries over, and the suite runs faster.

Validation: full Rust suite (1,087 passed, 19 ignored), clippy clean on both feature sets with the pinned 1.97 toolchain, 27 black-box E2E scenarios on the release build, 11 regression cases, the 10-case startup responsiveness suite, and the Python skill, gateway and ledger checks.