fix: restore model selection header + DSH/agent SSE stream termination & true tool streaming

[kirigayakazima]

Closes #82 — model selection is broken upstream: payload slot79 no longer routes models, every request is served by the account default.

Problem 1: Model routing is ignored

Google moved model routing from the payload slot79 field to the x-goog-ext-525001261-jspb request header (internal model ID selector). Without that header, slot79 is ignored and every model name in MODELS comes back as the same account-default model. Issue #82 documents this in detail.

Problem 2: Streams never terminate for strict OpenAI clients (DSH / Codex)

The SSE responses were sent over HTTP/1.1 without Content-Length and without Transfer-Encoding: chunked. Python's BaseHTTPRequestHandler does not emit chunked encoding automatically, so strict clients (OpenAI SDK, used by coding agents like DeepSeek Harness) wait forever for a body terminator that never arrives — the connection stays open (keep-alive), the stream never ends, and the client shows an endless thinking state until its idle timeout (300s) fires.

Fix: protocol_version = HTTP/1.0 → the connection closes after the SSE body (EOF), which is the canonical SSE termination signal.

Problem 3: Tool requests were not truly streaming

When tools were present, the handler fell back to a blocking full-generate-then-single-chunk path: it waited for Google to finish the entire response, parsed tool_calls, then returned one chunk. For agent clients sending 40+ tools, this produced multi-minute latencies and idle timeouts.

Fix: true streaming with tools — stream text chunks as they arrive (fast time-to-first-token), then emit the parsed tool_calls delta + finish_reason: tool_calls at the end.

Problem 4: Verbose tool descriptions blow up the prompt

Coding agents send 43 tools with long descriptions (~32KB JSON). Google Web takes minutes to process such a large prompt, again hitting client idle timeouts.

Fix: compact tool descriptions (trim each to ~150 chars, keep full parameter schemas) — ~32KB → ~12KB, responses in seconds while tools stay usable.

Changes

• MODEL_IDS map + build_model_header() emitting x-goog-ext-525001261-jspb (threaded through both stream and non-stream paths)
• fetch_xsrf_token() auto-discovery (FdrFJe, successor of SNlM0e)
• load_cookie() accepts gemini-auth.json and injects xsrf/gemini_bl/auth_user
• Add gemini-3.8-flash model entry
• HTTP/1.0 SSE termination (fixes endless thinking / idle timeout)
• True streaming with tools (fast TTFT + proper tool_calls delta)
• Tool description compaction (full parameters, trimmed descriptions)
• Real-time server.log output (stderr + file)

Verification

Tested with a paid AI Pro account + DeepSeek Harness (DSH) as the OpenAI-compatible client:

• Model routing confirmed via authoritative slot39 (56fdd199312815e2 served, not the cf41b0e0dd7d53e5 reject marker)
• Streaming terminates correctly: OpenAI SDK receives content chunks, then finish_reason, then [DONE] — no more endless thinking
• 43-tool agent requests complete in ~9s instead of timing out at 300s

Related

#88 (xsrf auto-refresh) and #91 (3.8-flash) are open PRs covering similar territory; happy to drop overlapping parts if those merge first.