fix(providers): PR #495 follow-ups — copilot 401 retry, real SSE streaming, device-flow errors, gemini env keys (#501)

[Milhouszhang]

Fixes the four defects found in post-merge live verification of PR #495. Closes #501.

Changes

• 74772b2e fix(openai): Copilot 401 now evicts the stale bearer, re-exchanges the GitHub OAuth token, and retries once (non-streaming + streaming); Chat Completions adapter now sends stream: true and parses real SSE — tolerates both Gemini-style complete tool-call chunks and OpenAI-style fragmented deltas
• cd717f78 fix(copilot): device-flow polling network/malformed/non-2xx failures now propagate as errors instead of Pending, so the desktop UI's consecutive-error path engages; authorization_pending/slow_down keep protocol semantics
• 25c2a064 fix(auth): non-interactive auth hydrates GEMINI_API_KEY (with GOOGLE_API_KEY fallback) for the google provider
• 5022e650 test(cli): tmux agent-loop tests set PUFFER_HOME so first-run onboarding doesn't intercept queued turns

Verification

• Unit/integration: mock-server tests for 401-retry-once and no-infinite-retry; SSE fixture tests for both Gemini and OpenAI streaming shapes; device-flow error-mapping tests; env hydration tests. cargo test -p puffer-core, focused puffer-cli suites, cargo check --all-targets, rustfmt/cargo fmt --check all pass. scripts/ci-gates.sh --quick passes (desktop frontend lane skipped: svelte-check not installed locally; changes are Rust-only).
• Live E2E on this branch's binary: real Gemini (/v1beta/openai/chat/completions) — env-var auth (both aliases), real SSE streaming (7 network-paced deltas), native function-call turn, zero errors; real GitHub Copilot — device-flow login end-to-end, token exchange, live streamed turn (50 SSE deltas + clean turn-complete). Copilot 401 mid-turn recovery remains mock-verified (bearer cache design prevents external injection of a real 401).
• Known pre-existing: tmux_snapshots failures exist on master (environmental, sandbox denies tmux sockets).

🤖 Generated with Claude Code