docs(CLAUDE): main-branch-everywhere policy + upstream provenance (tvm/tvm-ffi from apache)

Author: jeweldave

CLAUDE.md

@@ -19,6 +19,28 @@ fallback that produces a degraded/wrong/zero result is what is forbidden.
 This is the first and main rule. Existing automated fallbacks must be found and removed; route
 everything through the single clear path that fails loud.
 
+## Repos, branches, upstream provenance (David, 2026-05-30)
+
+All our repos live in the **DatasunriseOU** org and we work on the **`main`** branch EVERYWHERE
+(local + gb10 + every machine). Do NOT create side/feature branches for our own work — commit to
+`main`, push `main`, pull `main`. (Past mistake: scattering fixes across `m4-metal4-keystone` /
+`merge/upstream-codegen-reorg` side branches — never do this; consolidate on `main`.)
+
+Our repos (all in DatasunriseOU): `cppmega_mlx`, `tilelang`, `tvm`, `tvm-ffi`, `mlx`, `mlx-lm`,
+`cppmega`, `triton-pr9701`. These are OUR working repos (not "forks we don't control") — we commit
+to them directly.
+
+**Upstream provenance + how we pull fixes (we cherry-pick needed fixes FROM upstream INTO our main):**
+- **tvm** = `apache/tvm` upstream (git remote `origin` = https://github.com/apache/tvm). Our fork =
+  `DatasunriseOU/tvm` (remote `datasunrise`). We take tvm from **apache**, NOT from the TileLang/tvm
+  fork (`tilelang` remote = TileLang/tvm exists only to cherry-pick specific fixes).
+- **tvm-ffi** = `apache/tvm-ffi` upstream (Apache-2.0). Our fork = `DatasunriseOU/tvm-ffi`, vendored
+  as the `3rdparty/tvm-ffi` submodule inside our tvm. From **apache**, not tilelang.
+- **tilelang** = upstream `TileLang/tilelang`; our fork `DatasunriseOU/tilelang`, work on `main`.
+- **mlx / mlx-lm** = upstream `ml-explore/*`; our forks `DatasunriseOU/mlx`, `DatasunriseOU/mlx-lm`.
+- The pattern: keep an upstream remote (`origin`=apache for tvm, etc.), pull/cherry-pick the fixes
+  WE need from upstream into OUR `main`, and build from our `main`.
+
 ## Tensor memory rule (see AGENTS.md)
 - Wrappers/adapters must not silently allocate, cast, or copy large tensors to satisfy a
   boundary. If a design seems to require it, the design is wrong — keep looking (this is a