Check plugin output and post markup, fix post errors

[bits-bytes-nn]

Adversarial review rounds 1-16 plus a final round under the agreed criterion (defects in current posts or output, false positives on existing posts, defects from patterns existing posts use). Final round: 0 MAJOR.

_data/post_conventions.yml (scholar-lens copies this file)

• New key languages: [ko, en]: the allowed lang values; a translation is one post in each.
• Comments reworded (header; forbidden_keys note on last_modified_at).

New lint rules (script/lint-posts.rb), with pass / fail examples

Rule	Fails	Passes
$$ math needs use_math: true	식 $$x$$입니다. without the flag	the same with use_math: true; echo $$ in a fence
| inside $$ in a table row (kramdown splits the cell; | becomes |)	| $$p(y|x)$$ |, | $$|\theta|$$ |	| $$p(y \mid x)$$ |, | $$\lVert w \rVert$$ |
translation pair: one ko and one en post share translation_id; an -en post sets it	Korean original without translation_id	both posts with it
description: is plain text	$$, *emphasis*, backticks, ](, <tag>, {{	Mixtral 8*7B와 C++ [1]
Liquid outside {% raw %}	```text\nQuestion: {{ question }}\n```	the same inside {% raw %}…{% endraw %}; render_with_liquid: false
tags read as HTML	<think>, <s>, <input>, bare <img> in prose	`<think>`, \<think\>, <br>, <img src=…>, i<j
code fence kramdown does not read	```python title="a.py"; a fence never closed	```python
file name YYYY-MM-DD-; date: not after it (UTC)	date: 2025-03-01 in 2025-02-27-*.md	a paper's earlier date
tag spelled two ways across posts	Sliding-window-Attention next to Sliding-Window-Attention	one spelling
only .md files under _posts/; front matter on line 1	x.mkd; a blank line before ---	—

Other gates

• script/check-plugin-output.rb (run by validate-site.sh): reads the posts through Jekyll, checks each hook's mark (legacy redirect stub, modified time, prev/next, image loading/size, table container), the feed cap, and block HTML kramdown could not parse (closing tag shipped as text).
• validate-site.sh: every Ruby check fails when it crashes; listing check scans every non-post page; hreflang-gone guard; site author: required; external redirect_to and directory/bare permalinks handled.
• _config.yml: liquid.strict_filters: true; future: true (a same-day post dated 12:00 UTC no longer drops out of a morning deploy).
• CI htmlproofer no longer passes --no-enforce-https; all arXiv links are https.

Hand fixes to generated posts (source value vs. what the post said)

• Llama 3 405B: MMLU 87.3, GSM8K 96.8 (Table 2); post said 85.2 / 89.0 (fixed in Add descriptions to paper posts and fix factual errors #88, listed for completeness).
• Gemini 2.5: SFT data "loosely inspired by Constitutional AI" plus human revision; RL reward from a reward model and a rubric Critic; post said the model critiques itself via Constitutional AI. Video and benchmark figures attributed to 2.5 Pro.
• DeepSeek-V3: MMLU-Pro 75.9 / GPQA 59.1 best among open models; MATH-500 90.2 best among compared models; post said best overall.
• LightRAG: GraphRAG retrieval needs "hundreds of" API calls, 610 × 1,000 / C_max; post said 610.
• Gemma 2: 0.00026% is the share of memorized outputs with low-severity personal data (post: "emission rate"); the 15T-token point cites Llama 3, not Chinchilla (two places).
• Gemma 3: Gemma 3 27B IT tied 9th (Elo 1338); Llama 3.1 405B 1269 (post said 1257, Qwen2.5-72B's score); GPT-4.5-Preview tied 1st.
• Qwen3: the 1/3 total, 2/3 active comparison is Qwen3-235B-A22B-Base vs DeepSeek-V3-Base, ahead on 14 of 15.
• DeepSeek-V2: AlpacaEval 38.9 / MT-Bench 8.97 are DeepSeek-V2 Chat (RL).
• Zep: LongMemEval 15.2% / 18.5% are relative gains (55.4→63.8, 60.2→71.2); post said percentage points.
• LLaMA, DeepSeek-Coder, Mixtral, Mistral 7B: "most benchmarks" / "almost all" where the post's own tables show losses.
• RAPTOR: +20 percentage points on QuALITY. GraphRAG: "over 97%" fewer tokens.
• DeepSeek-R1: <think>/<answer> in the prompt-template table were swallowed as HTML; now code spans.

Not changed (pending a decision)

Paper posts publish date: = paper date as datePublished / feed published.

Gates: tests (169 runs), post lint, strict build, htmlproofer (CI flags), validate-site, shellcheck, actionlint; MathJax 0 errors on all 30 math posts.