fix: Gemini turn ordering and SSE streaming parser

[boxcee]

Three bugs fixed in the Gemini driver:

1. Function call turn ordering: Gemini requires model turns with functionCall to be immediately followed by user turns with functionResponse. The agent loop could insert text-only turns between them (e.g. "[no response]", "Please continue"), causing INVALID_ARGUMENT 400 errors.

2. First turn must be user: Gemini rejects conversations starting with a model turn. After session trimming or compaction, the first message could be a model turn with functionCall parts. Now prepends a synthetic user turn when needed.

3. SSE streaming parser: The parser used \n\n as the SSE event delimiter but Gemini returns \r\n\r\n (HTTP standard). Since \r\n\r\n does not contain the substring \n\n, no events were ever parsed, causing 0 token responses and crash loops. Fixed by normalizing \r\n to \n before delimiter matching.

Also adds debug logging for turn structure, request/response bodies, and SSE stream diagnostics.

Summary

Changes

Testing

• cargo clippy --workspace --all-targets -- -D warnings passes
• cargo test --workspace passes
• Live integration tested (if applicable)

Security

• No new unsafe code
• No secrets or API keys in diff
• User input validated at boundaries

[jaberjaber23]

The bugs being fixed are real and important — the SSE \r\n issue and turn ordering constraint are genuine Gemini API problems.

Blocking issues:

1. truncate_for_log will panic on multi-byte UTF-8 input. &s[..max_len] indexes by byte offset — if it falls in the middle of a multi-byte character (Japanese, emoji, non-ASCII error messages from Google), Rust panics. Fix with s.floor_char_boundary(max_len) (stable since 1.76).

2. buffer.replace("\r\n", "\n") runs on the entire accumulated buffer every chunk, creating O(n*m) behavior. Move normalization to the chunk level: buffer.push_str(&chunk_str.replace("\r\n", "\n")) instead of appending then replacing the full buffer.

3. No tests for the 70-line enforce_function_call_ordering function. This is critical LLM infrastructure — needs tests covering: functionCall with intervening text, orphaned functionCall, conversation starting with model turn, consecutive same-role merging.