feat: expose resolved renderer prefix stability

[hallerite]

Expose renderer.is_prefix_stable after renderer configuration and chat-template kwargs are resolved. The flag describes whether full renders ending in an assistant remain token prefixes when messages are appended, with a fixed preamble and tools. Unknown custom templates are conservatively false.

Configuration-sensitive renderers account for history reasoning retention, last-assistant wrappers, effort hints, and training mode. DeepSeek V4 and Inkling are conservatively false because supported message controls and reused tool-call IDs can rewrite earlier output.

A manual, opt-in audit uses tests/fixtures/prefix_stability.json: 36 executable instability witnesses and 21 stable controls sharing six conversations. Each case stores the renderer/config, messages to append, and an explanation. Tests compare actual token prefixes and show the first differing tokens plus a decoded diff on failure. Coverage checks include every registered built-in renderer; opaque DefaultRenderer is tested as unknown. Finite stable controls are not a proof for arbitrary inputs.

Run explicitly with uv run pytest tests/audit_prefix_stability.py -q. The audit is excluded from normal pytest discovery and CI, and reuses the existing shared tokenizer cache.

Validation: manual audit 59 passed in 9.80s; collection-only verification confirms the normal suite excludes the audit. Ruff lint/format and git diff --check passed. The full suite was not rerun locally.

Closes #41.

Rebased onto origin/main at 376f54f (Python 3.11 typing update). The focused manual audit and Ruff checks pass after the rebase; all four PR commits are signed.

Note

Add is_prefix_stable property to the renderer protocol and all built-in renderers

• Adds an is_prefix_stable boolean property to the runtime-checkable Renderer protocol in base.py and implements it in every built-in renderer, classifying whether a full re-render preserves a completed conversation prefix.
• Some renderers return a constant value (e.g. DeepSeekV3Renderer is stable; Qwen3Renderer is not). Others derive the value from configuration, such as GLM5Renderer (clear_thinking, enable_thinking), Qwen35Renderer (preserve_thinking), Hy3Renderer (raw_last_assistant, is_training), and Nemotron3Renderer (truncate_history_thinking, effort hint).
• Adds an opt-in prefix-stability audit in audit_prefix_stability.py driven by a witness corpus in prefix_stability.json. The audit renders original and appended conversations, checks each renderer's reported value against the corpus, and fails if any registered built-in renderer lacks a corpus case. Failures include token-level and decoded-text prefix diagnostics.
• Documents the guarantee, known stable/unstable cases, and the audit workflow in README.md. SFT behavior is unchanged and still produces one sample per conversation.
• Risk: third-party or custom renderers that do not define is_prefix_stable will fail the runtime protocol check if it is used for isinstance-style validation.

^{Macroscope summarized 0e4694a.}

---

Note

Low Risk
Additive API and documentation; behavior of render() and bridges is unchanged. Custom renderers without the property may need getattr if consumers use strict protocol checks.

Overview
Adds renderer.is_prefix_stable to the Renderer protocol so callers (especially SFT) can tell whether a full re-render of a conversation that ends on an assistant keeps the same token prefix when more turns are appended (add_generation_prompt=False, fixed tools/config). DefaultRenderer always reports False; each hand-coded renderer sets a constant or derives the flag from template knobs (e.g. GLM-5 clear_thinking, Qwen3.6/3.8 preserve_thinking, Hy3 training/preserved thinking, Nemotron effort/truncation).

README documents the guarantee, how it differs from thinking_retention / bridge_to_next_turn, and conservative getattr(..., False) for custom renderers.

Opt-in audit (tests/audit_prefix_stability.py + tests/fixtures/prefix_stability.json): parametrized witnesses compare token prefixes before/after append and assert they match each renderer’s declared stability; includes registry coverage and a Qwen3.8 chat_template_kwargs check. Not run in default CI — manual uv run pytest tests/audit_prefix_stability.py -q.

^{Reviewed by Cursor Bugbot for commit 0e4694a. Bugbot is set up for automated code reviews on this repo. Configure here.}

[macroscopeapp]

Approvability

Verdict: Approved at 6d8c744

Macroscope's review found this PR approvable — This is an additive renderer metadata contract with configuration-aware boolean properties; existing rendering, bridging, and SFT behavior remain unchanged. Documentation and an opt-in audit corpus cover the new contract, with only the documented need for custom implementations to expose the property as a compatibility consideration.

Notes:

• Macroscope's correctness review did not run, so approvability was decided on eligibility alone.

No code changes detected at 0e4694a. Prior analysis still applies.

^{You can add or adjust custom eligibility rules. Learn more.}