v3.23.1 — a 6% gateway blip no longer kills the whole session

Telemetry (audited 2026-05-28) showed the Solana gateway intermittently returns PaymentRejected on ~6% of paid-model calls (28/468) — identical prompts succeed five seconds apart. Three client-side defects turned that transient blip into "session ruined, and restart doesn't help." This release caps the blast radius at a single turn:

• payment_rejected is now transient. The classifier retries it up to 3× with backoff instead of surfacing a hard error on the first blip. Each retry signs a fresh nonce, so a retry is not a replay — genuinely broken wallets (wrong chain, clock skew) still exhaust the small budget quickly and see the same guidance.
• One blip no longer demotes you for the whole session. payment_rejected now falls back to a free model for that turn only — it isn't added to the session-permanent payment blacklist and carries no elo penalty, and the next turn resets to your chosen model. (Genuine 402s / insufficient funds stay session-permanent, as before.) The per-turn fallback also resets the retry counter so the free model gets its own budget rather than inheriting the exhausted one.
• /exit no longer leaves a zombie. MCP shutdown is now a bounded 2s race followed by an explicit process.exit() in both the Ink and basic UIs, so keep-alive sockets and MCP child processes can't pin the event loop after "Goodbye."

The gateway-side root cause (a Solana nonce-cache race) is tracked separately; this release is the client-side blast-radius cap.

435/435 local tests pass; e2e green against live models.

Full PR: #70