fix(gemini): emit reasoning chunks and fix thought parts extraction

[kirillreutski]

Problem

When using GeminiProvider with thinking-capable models (e.g. gemini-2.5-pro), reasoning content is silently lost in three separate places:

1. includeThoughts missing from request — Gemini API requires includeThoughts: true in thinkingConfig, otherwise the model thinks internally but returns no thought parts in the response. The SDK was setting thinkingBudget/thinkingLevel without this flag.

2. stream() drops thought parts — parseGeminiChunk lumped all text parts into textChunks without distinguishing part.thought === true. Thought content was emitted as regular text instead of think_chunk_* SSE events.

3. collectAll path drops thought parts — The fallback path (when Gemini returns a JSON array instead of line-by-line SSE) had the same problem: thoughtChunks were never extracted, so the agent never emitted reasoning events on this path either.

A fourth issue: empty part.text strings ("") were being pushed into textChunks, causing spurious empty text_chunk SSE events.

Changes

src/infra/providers/gemini.ts

• _buildGenerationConfig: add includeThoughts: true whenever thinkingBudget or thinkingLevel is set.

• parseGeminiChunk: split parts by part.thought === true into a new thoughtChunks: string[] return field (only when reasoningTransport === 'provider'); skip empty strings.

• stream() — SSE path: consume thoughtChunks and yield content_block_start(reasoning) + content_block_delta(reasoning_delta) events before text chunks.

• stream() — collectAll path: same reasoning-chunk handling for the JSON-array fallback path.

• parseBlocks (used by complete()): map part.thought === true parts to { type: 'reasoning', reasoning } blocks instead of text blocks.

Testing

Verified end-to-end with gemini-2.5-pro using reasoningTransport: 'provider':

• think_chunk_start / think_chunk / think_chunk_end SSE events now arrive correctly in streaming mode
• complete() returns reasoning blocks in the message content
• No regressions on non-thinking models (thought parts absent → thoughtChunks stays empty)

[CrazyBoyM]

Thank you @kirillreutski for identifying the Gemini thought-loss paths and putting together the original fix.

I opened #58 to carry this through the repository's full merge gate. Rather than cherry-picking or rewriting your work, #58 merges all four commits from this PR unchanged: 51acf76, 4ddd074, 5f1f665, and ff5cf2a remain in its ancestry with your original authorship.

The follow-up adds the protocol hardening and deterministic coverage found during pre-merge review: complete block lifecycles, ordered interleaved parts, all history transports, thought signatures (including signature-only empty parts), function-call IDs, robust SSE/JSON fallback parsing, Agent persistence/privacy behavior, and Node 20/full-suite checks.

I have requested your review on #58 as a courtesy; no action is required for attribution. Once #58 passes the full CI chain and is merged with a merge commit, this PR can be closed as incorporated/superseded without losing its history.