README: make clear this is a fork, attribute upstream content explicitly

[dkurukula]

Problem

Every claim, benchmark number, vision statement, and roadmap item in the README was written in upstream's (JustVugg/colibri) voice with no indication that JustVugg/colibri and dkurukula/colibri are different things. A reader landing on this fork's README had no way to tell upstream's 6x RTX 5090 numbers from this fork's actual hardware, or upstream's stated roadmap from a commitment this fork is making. Requested directly: make it clear what's this fork's and what's upstream's, and attribute every instance of upstream-maintainer voice explicitly.

Fix

Added a "This fork" section up front, right after the badges, stating plainly:

• this is dkurukula/colibri, a persistent fork of JustVugg/colibri, not a snapshot or one-off patch;
• the actual hardware it's tuned for: a Dell PowerEdge R720, dual Intel Xeon E5-2660 v2 ("Ivy Bridge," no AVX2/FMA), 125GB RAM, one NVIDIA GTX 1050 Ti (4GB VRAM) — pulled from /proc/cpuinfo and nvidia-smi on the actual box, not guessed;
• an accurate list of the 14 merged fork-specific PRs, built from the real GitHub PR history (gh/API query against dkurukula/colibri), not summarized from memory: the AVX-only Sandy/Ivy Bridge kernel paths (without which this fork's CPU doesn't even build the fast path), CUDA_MISS_GPU, its async double-buffered dispatch pipeline, the COLI_MISS_CPU_EVERY CPU/GPU concurrent split (inspired by FreeToken, arXiv:2608.16157, re-derived against this hardware rather than copied), the ported ATSInfer live re-pin gate (arXiv:2607.10183 Algorithm 3), the decode heartbeat, stall watchdog, systemd deploy templates, live-usage watcher, and the podman quick-start path.

Every remaining section that is upstream's own voice — the vision statement, the "What it achieves" benchmark numbers, the roadmap, the community/sponsorship ask, the naming rationale, the acknowledgements — is now explicitly marked with an attribution line ("Upstream's own stated vision:", etc.) and mostly converted to a blockquote, so it reads as a quote rather than this fork's own claim. Upstream issue links (JustVugg/colibri/issues/NNN) are labeled "(upstream issue)" so they don't read as this repo's own tracker. The website/release badges at the top are relabeled "upstream" explicitly, and a new badge/link to the upstream repo itself was added.

Added this fork's own real, measured numbers for honest comparison: decode at 0.123 tok/s with CUDA_MISS_GPU alone, 0.156 tok/s with this fork's CPU/GPU split also enabled (~27% improvement, measured on the live production deployment, cap=3, single GTX 1050 Ti — see the cuda-miss-cpu-gpu-split PR history for full methodology). Stated plainly that this lands close to upstream's own documented "25GB dev box" floor (0.05-0.1 tok/s), nowhere near their flagship 6-GPU numbers — this fork's hardware is old and constrained by upstream's standards, and the README now says so instead of leaving upstream's benchmark table looking like this fork's own results.

"Get started" clone URLs now point at this fork (dkurukula/colibri) rather than upstream, since that's what running this specific README implies; the prebuilt-release path is kept but explicitly marked as upstream's own releases (this fork doesn't publish separate binaries).

Also fixed while in there

c/glm.c → c/colibri.c in two places (the single-file engine reference and the repo-layout tree) — stale even relative to upstream itself, which renamed the file per their own CHANGELOG JustVugg#391; the README just never caught up. Small, adjacent, factual correction, not scope creep.

Testing

Docs-only change. Verified: code fences balanced, all 13 section headers intact, no broken markdown structure. No build/engine files touched.

---

Generated by Claude Code