feat(laguna): add poolside/Laguna-S-2.1 support via a dedicated renderer

[eexwhyzee]

Summary

Adds renderer support for poolside/Laguna-S-2.1, a larger sibling of XS-2.1.

S-2.1 shares XS-2.1's tokenizer (same vocab, merges, and special tokens/ids) but ships a different chat template: it defaults enable_thinking to True (XS-2.1 defaults False) and adds a preserve_thinking kwarg that widens the reasoning-display gate to enable_thinking or preserve_thinking. The token format is otherwise identical.

Note: an earlier version of this PR routed S-2.1 to the existing XS-2.1 renderer on the assumption the chat template was byte-identical. It isn't — that mismatched apply_chat_template on the default thinking flag (bare </think> vs an empty <think></think> wrapper) and failed test_render_ids / test_build_helpers. (The config-parity barrage missed it because it always sets enable_thinking explicitly, never exercising the differing default.) This PR now gives S-2.1 its own renderer/config.

Changes

• LagunaS21RendererConfig — enable_thinking=True, preserve_thinking=False defaults, matching S-2.1's template. Registered in the discriminated union, _CONFIG_BY_NAME, and exports.
• LagunaS21Renderer — a thin LagunaXS21Renderer subclass that overrides only the reasoning-display gate (enable_thinking or preserve_thinking); everything else (role tags, tool-call packing, system-block gate, generation prompt, parse skeleton) is inherited unchanged.
• Route poolside/Laguna-S-2.1 → laguna-s-2.1 in MODEL_RENDERER_MAP; register the renderer; lazy-import + exports.
• Typing follows the existing GLM5/Nemotron subclass idiom (base __init__ accepts the config union; subclass narrows self.config).

Testing

• Full suite: 2700 passed, 128 skipped, 1 xfailed, 0 failures.
• New tests/test_laguna_s21.py: token-exact parity vs apply_chat_template for all four enable_thinking × preserve_thinking combinations, plus the thinking-on default and the preserve_thinking history-retention behaviour.
• S-2.1 added to the shared render-parity matrices (config-parity, is_content, sampled_mask, render_ids, build_helpers) — all green.
• ruff + ty clean (no new diagnostics vs main).

🤖 Generated with Claude Code

[macroscopeapp]

Approvability

Verdict: Needs human review

This PR adds a new model renderer (LagunaS21Renderer) for poolside/Laguna-S-2.1, introducing new user-facing capability. While the implementation is a thin subclass following existing patterns, new feature additions warrant human review to verify the integration and default behavior differences.

^{You can customize Macroscope's approvability policy. Learn more.}