feat(config): securely configure local Cartesia and OpenAI keys

[axelpey]

Owners can now enter Cartesia and OpenAI keys through innate keys set cartesia / innate keys set openai, inspect configuration presence with innate keys status, and remove keys without putting credentials into browser settings, ROS parameters, command arguments, or output. Entry uses a hidden terminal prompt or a secret-manager pipe via --stdin; changes take effect after a node/local-simulator restart.

The CLI reuses the simulator setup's .env persistence path. Updates are atomic, locked across processes, and mode 0600; replacing a key removes stale commented copies. Removal writes an empty override so old shell/system credentials do not silently return on restart. Generated simulator env files are also private and retain empty overrides. Local .env values now consistently take precedence over inherited shell keys, matching the robot launch loader. INNATE_PUBLIC_DEMO=1 rejects key commands and credential-bearing generated environments.

Runtime dependencies and boundaries

• Direct Cartesia TTS already belongs to David's feat(tts): let CARTESIA_API_KEY stand in for a service key [blocked: Alfred voice not public] #755. This PR does not duplicate its transport, change his branch/status, replace Alfred, or claim to resolve his external voice-access blocker.
• The complete OpenAI Responses agent/provider selection is in feat(brain): experimental model and thinking cadence foundation #772. The safe transport contributed from this work is included there, not duplicated in this diff. The official model ID is gpt-6-astra; entering a key does not select a model.
• Configured service-key routing is preserved. No deployed robot/public simulator was modified. The public-demo flag is a deployment guard; runtime credential isolation remains the separate security audit's relay work.

Verification

50 focused tests passed: actual CLI subprocess save/rotate/status/remove, real PTY hidden input and cancellation, actual hardware .env loader and simulator generated environment, service-key preservation, shell override removal, public-demo rejection, concurrent process writes, atomic failure cleanup, permissions, symlink refusal, and launcher imports. Ruff and diff checks pass. Latest head 941d678f4 passes GitHub format and host pytest; ROS integration/image builds are still pending.

Combined verification against #755 plus #774 (8e2d0d76a) and #772 (1cc44b47c) also passed: actual CLI pipe → private .env → hardware launch loader → OpenAIContext two-turn native tool replay and Cartesia TTSHandler → local HTTP provider doubles, plus 54 provider/TTS tests. The integration needed two merge resolutions with #755: retain all four SECRET_ENV_KEYS in sim/launcher/config.py, and retain David's generic setup UX while importing the extracted env-store helpers in setup_wizard.py. No foundation conflict occurred.

The OpenAI transport's additional 8 real local HTTP tests passed before handoff to #772 (direct/proxy routing, no account failover, malformed/incomplete/error stream sanitization, missing key, public-demo guard). Live OpenAI account/model authorization and Cartesia synthesis with a non-owning key are not verified in this PR. Those require an owner-provided key through the local secret path; never paste credentials into a PR or chat.

Keep this PR draft until Axel explicitly approves it.