docs: rewrite README and add the developer guide

[bowencui123]

The documentation half of PR #312. Stacked on #313, so the diff below is only the 11 documentation and asset files; retarget the base to main once #313 lands.

What this does

• README rewritten around the results a reader wants first: the pairwise win/loss matrix, the suite geometric means, the sweep-max latency figure and the overview diagram, then the suite, backends, installation, quick start, methodology and the LLM track.
• docs/developer_guide.md added: the CLI reference plus the former OPERATOR_AUTHORING_GUIDE.md, which is removed from the repository root.
• assets/: the TileBench logo as the page title (cropped to its content; a transparent variant for light mode and the original cream card for dark mode, so the black wordmark stays readable), the overview diagram with a transparent background, the sweep-max latency figure, and four link icons. Link icons appear only in the top resource bar and the Recorded Results table; inline body links are plain Markdown. Each ## section heading carries an emoji, in the style of the SWE-bench README, and the table-of-contents links use the matching #-name anchors.
• Artifact access. The README links the three places results live outside main: the 220 raw Nsight Compute reports on Hugging Face (free login required), the raw logs on archive/raw-logs-2026-09-18, and the LLM campaign on Google Drive. It states that tilebench/benchmarks/llm_generated/ is the Git-ignored runtime output path, that a fresh clone does not ship the 6,140 campaign files, and that python scripts/fetch_artifacts.py --artifact llm-aacl2026 restores them.
• Hardware-scoped results. Follows item 9 of chore: repository cleanup and tilebench package layout #313. Every command passes --gpu B200, and result paths read results/<hardware>/{csv,logs,figures,aggregate,runs}/. The text says that this is a general layout, that results/B200/csv/ holds the paper results committed today, and that other hardware sits beside it without a code change. The developer guide documents the label rules, what a namespace holds, the scope of the TileLang merge, and the per-GPU archive command. Both documents state that NKI results are stored with the B200 campaign as cross-hardware measurements taken on AWS Trainium: three columns in results/B200/csv/, logs and profiles under results/B200/logs/, with speedup_nki = torch_nki_ms / nki_ms and never torch_ms / nki_ms. The headline B200 numbers are unchanged, and no empty GH200 or AMD directory is created.
• Multi-hardware consistency. Follows items 10 to 13 of chore: repository cleanup and tilebench package layout #313. Profiling metadata is documented as tilebench/profiling/metadata/<hardware>/{ncu_catalogue,kernel_counts}.json, with B200 as the directory committed today, and NCU reports as outputs/ncu/<hardware>/. The 220 released reports are described as the paper's B200 profiles; the count is unchanged. Raw JSON names carry the mode and the backend selection: the README shows the two files its quick start produces and passes --tile-language to visualize.py, and the developer guide adds the naming rule, --mode, the ncu_catalogue log contract and the --gpu commands of the NCU tools. The paper's B200 raw logs are stated to live under results/B200/logs/ on the archive branch; the sentence that described results/logs/ as their current location is gone.

Verification

• The win/loss matrix (37/8, 36/9, 33/12) and the "14 operators within 5%" claim recompute exactly from results/B200/csv/; the geometric means (2.02×, 1.58×, 1.18×, 1.22×), the suite composition (26 from TritonBench, 19 from LeetGPU) and the category counts match the paper sources.
• Every operator sweeps 20 cases per dtype, as stated.
• All 22 CLI flags cited across both documents exist in the corresponding --help; the commands no longer need PYTHONPATH= and were run as written. The flags added for the results layout (--gpu on four scripts, --output on plot_sweep_max.py, --results-root) were checked against --help as well, and the README quick-start form of run_bench.py was run for real on a B200 under a throwaway label.
• On GitHub's render of this branch all 38 in-page links resolve, the logo, both figures and the seven link icons load, and the cross-links between README and the developer guide work both ways.
• Every path reflects the layout from chore: repository cleanup and tilebench package layout #313; no pre-refactor path remains.
• The documented fetch command was run for real against the published Google Drive file: download, SHA256 match, extraction, and a tree comparison that reproduced the 6,140-file campaign exactly.
• The four external targets (Hugging Face dataset, Google Drive file, archive branch, developer guide) return HTTP 200. pytest on this branch: 183 passed, 1 skipped.

Two things to decide before merge

1. Decided: the backend table keeps TileLang "Results: N/A". The table is unchanged in this PR. For reference, all 45 committed CSVs under results/B200/csv/ carry tilelang_ms and speedup_tilelang (suite geometric mean 1.72× over PyTorch); they are not part of the headline comparison.
2. The citation section and BibTeX entry were dropped in the rewrite. Intentional while the proceedings entry is pending, or should they come back?

Separately, the repository still has no LICENSE file while the README tells people to pip install -e ..

🤖 Generated with Claude Code

https://claude.ai/code/session_01EdNBRsQvET5PUJ3Nw1hVCA