feat: add all 7 Qwen3.5 sizes to MODEL_RENDERER_MAP, autodetect think polarity

[hallerite]

Summary

• Adds all seven Qwen3.5 sizes to MODEL_RENDERER_MAP: 0.8B, 2B, 4B, 9B, 35B-A3B, 122B-A10B, 397B-A17B (previously only the middle three were listed).
• Teaches Qwen35Renderer to autodetect the enable_thinking default from the tokenizer's chat template at construction.
• Bumps version 0.1.7 → 0.1.8.

Why

Qwen3.5 ships two chat-template variants that differ only in the polarity of the gen-prompt branch:

Variant	Sizes	Default at gen prompt
Big-default-on	4B / 9B / 35B-A3B / 122B-A10B / 397B-A17B	<think>\n (thinking on)
Small-default-off	0.8B / 2B	<think>\n\n</think>\n\n (empty block)

The previous Qwen35Renderer hardcoded enable_thinking=True, which matched the big-default variant but not the small one — so 0.8B / 2B couldn't be added to the map without silently breaking byte parity at add_generation_prompt=True.

Implementation

Qwen35Renderer.__init__(... enable_thinking: bool | None = None):

• When None (the new default), runs a one-shot tokenizer.apply_chat_template(...) probe and reads the polarity off the resulting tail (</think> ⇒ default off).
• Explicit enable_thinking=True/False still wins — autodetect only fires when the caller omits the flag.
• Probe failures fall back to True (matches the big-default majority).

Tests

New tests/test_qwen35_size_coverage.py (37 cases):

1. Map shape — exactly the seven listed sizes route to qwen3.5; catches silent additions.
2. Polarity autodetect — each of the 7 sizes resolves to its expected default (big → True, small → False) when constructed without an explicit flag.
3. Byte parity barrage — for each of the 7 sizes × 4 representative conversations (system+user, single-turn, gen-prompt, with reasoning), Qwen35Renderer.render_ids matches apply_chat_template token-for-token.

Full suite: 937 passed, 48 skipped, 1 xfailed.

Test plan

• pytest tests/test_qwen35_size_coverage.py — 37 cases pass
• Full suite (pytest tests/ --ignore=tests/test_client.py) — 937 pass, no regressions
• Pre-commit hooks (ruff check + format) clean

🤖 Generated with Claude Code