Changelog

Adds android_log calls at each step of the unified recovery setup flow

(mnemonic generation, vault recovery, keypair generation, S3 upload) to

aid debugging on Android where stdout/stderr are not visible. Converts

existing eprintln calls to android_log for consistency.

SaveAiApiKeyAsync was calling RequestVaultUnlockAsync (which internally

initializes the vault if needed) and then calling VaultInitialize again

with a hardcoded password. The second call always failed with

AlreadyInitialized since the vault was just initialized by the unlock

flow. Simplified to a single VaultIsUnlocked + RequestVaultUnlockAsync

check which handles both initialization and unlock.

Replace StatelessExecutor with InteractiveExecutor using a fresh context

per inference request. StatelessExecutor was creating and destroying

multiple Metal contexts per call, potentially causing inference issues.

InteractiveExecutor uses a single stable context. Also added pre-

tokenization diagnostic logging (prompt token count) and set TokensUsed

in the response so AiService logs accurate token counts.

StatelessExecutor automatically prepends a BOS token, but the Llama 3.x

chat template also included <|begin_of_text|> (the BOS token), causing

a double BOS that confused the model into immediately emitting EOS and

returning 0 tokens. Removed the explicit BOS from the template.

Also changed GpuLayerCount from 0 (CPU-only) to -1 (offload all layers)

so the model uses Metal acceleration on Apple Silicon, which should

significantly improve inference speed.

The local LLama provider was hardcoded to use the Phi-3 chat template

(<|system|>...<|end|>) for all models. When running Llama 3.2 or

Mistral models, the unrecognized special tokens caused the model to

immediately emit EOS, returning 0 tokens. Added FormatPrompt() that

selects the correct chat template and anti-prompts per model family:

• Llama 3.x: <|begin_of_text|><|start_header_id|>... format
• Mistral: [INST]...[/INST] format
• Phi-3: <|system|>...<|end|> format (default)