policy-catalogue: add ggml-org/llama.cpp entry

[sjh9714]

Summary

Adds policy-catalogue/llama-cpp.md for ggml-org/llama.cpp, and updates the catalogue README (12 → 13 entries). Requested in #20.

llama.cpp adopted its current AI usage policy in PR #26012 (merged 2026-07-22), tidied in PR #26030 (2026-07-23). All links and verbatim excerpts are pinned to commit 555881e so the quotes survive link rot.

Three things in it that read as v0.2 signal rather than just another row:

1. First loosening in the catalogue. Every other entry restricts, holds, or closes; this one moved from a ban on fully/predominantly AI-generated PRs to allowing them, and the maintainer's stated rationale is improved model quality. That makes stance a time-varying property tied to model capability, which makes captured_on load-bearing rather than bookkeeping.
2. The split is artifact-vs-conversation, not code-vs-docs. Code is allowed; PR descriptions, commit messages, reviewer responses, bug reports, feature requests and discussion posts are all prohibited. The schema's asset dimension can't express that, and a review-interaction field alone wouldn't either — the prohibited set spans the whole contribution conversation.
3. A clause addressed to the agent's configuration. Like Godot the policy speaks to the agent, but this one instructs it to update its own memory/config to exclude the project — a directive aimed at the control plane, and unverifiable from the contribution artifact.

AI disclosure (required — we dogfood our own standard)

• AI-assisted — tool and role: drafted by Claude (Claude Code) working under my direction. Every verbatim excerpt, date, PR number and schema-field mapping was verified programmatically against the pinned upstream commit (555881e) before submission; the comparative claims about other catalogue entries were checked against the files in this repository.

Attestations (required)

• understands-change — I can explain every part of this entry and the sources behind it.
• reviewed-output — reviewed before submission. To be precise about the mechanism rather than just ticking the box: I reviewed a structured walkthrough of the entry plus the verification output confirming each quoted string exists at the pinned commit, rather than reading the rendered diff line-by-line myself. Flagging that explicitly because "what does review actually mean when an agent drafted it" seems like the live question for this standard, and a vague tick would be the wrong contribution to make here.
• ran-tests — N/A (prose only).

Type of contribution

• Policy-catalogue entry (single file in policy-catalogue/)

Sources

• AGENTS.md @ 555881e — https://github.com/ggml-org/llama.cpp/blob/555881ebc8b0fc0402b30e09258a32a7bfd13c52/AGENTS.md
• CONTRIBUTING.md § AI Usage Policy @ 555881e — https://github.com/ggml-org/llama.cpp/blob/555881ebc8b0fc0402b30e09258a32a7bfd13c52/CONTRIBUTING.md
• Pull request template @ 555881e — https://github.com/ggml-org/llama.cpp/blob/555881ebc8b0fc0402b30e09258a32a7bfd13c52/.github/pull_request_template.md
• Policy change — contrib: allow all AI-generated code in general ggml-org/llama.cpp#26012 (merged 2026-07-22)
• Follow-up cleanup — contrib: fix leftovers from the AI usage policy update ggml-org/llama.cpp#26030 (merged 2026-07-23)

Maintainership disclosure: I maintain MergeWarden, a checker in the layer-3 space discussed in #20. This entry documents llama.cpp's policy only and makes no claim about any tool.