[jayleaton]

Summary

Every engine-level change LocalRouter carries against the TensorFold engines now has an upstream pull request, or a recorded reason why it has none. docs/dev/UPSTREAM.md is the standing list. A new rule in CONTRIBUTING.md says any new engine-level patch gets an entry there and an upstream PR.

docs/dev/UPSTREAM.md   new: the audit and its status table (11 entries)
CONTRIBUTING.md        + "Upstream engine changes": entry + upstream PR for every new engine-level change

Change we carry	Upstream PR	Tested
Drama 3 voice prefix cache, CUDA full model (patches/engine/)	ashhart/TensorFold#1745	Host tests over a fake lane. Linux CUDA host and macOS Metal builds pass. No CUDA run (none here); the PR cites our DGX-01 numbers for the earlier patch.
Drama 3 voice prefix cache, CUDA NVFP4 (patches/engine-nvfp4/)	already proposed by someone else: plotarmordev/TensorFold#4	none opened (it would duplicate #4)
Launchers: --external-admission, plus --reserve-gib on the Linux launcher	plotarmordev/drama-tensorfold#35	Unit and macOS smoke tests, and the real NVFP4 Metal engine on this Mac: main refused at 47.7 GiB free, the branch was ready in 5.6 s and served speech.
macOS launcher: --parent-pid	plotarmordev/drama-tensorfold#36	Smoke tests with a real orphaning parent, and the real engine: it stopped cleanly when its supervisor exited.
Launchers under python3 -I	plotarmordev/drama-tensorfold#37	Fails on main, passes on the branch; the real engine ran under -I.
Streamed reasoning_content trailing newline (a qualification exception)	ashhart/TensorFold#1743	Qwen3.8-27B on Metal on this Mac: the stream went from 599 to 598 bytes (and 352 to 351), equal to the whole reply.
graph.endCapture leaves the capture scope open (caller-side captureEnd)	ashhart/TensorFold#1744	A fake-driver test fails on main. Linux build passes. No CUDA run.
CLIP Lowercase normalizer	not opened: no upstream model uses it (dead code by their rules)	none
Chat pinned to 1.1.1 on GB10	none: a pin choice, no code	none
cuBLASLt and driver identity (G1/G3)	not upstreamable: our attestation policy	none
NVFP4 kernels from v0.6.1	not upstreamable: unmodified copies of a frozen line	none

Upstream's own series, pinned by hash in pins.json (59 patches) and pins-macos.json (5), are drama-tensorfold's own files. All 64 sha256s match upstream, so they are skipped, as is apps/mobile's react-native patch.

Evidence

• Before: no list of what we carry against upstream, and no upstream PR for any of it.
  After: six upstream PRs open and linked above, and each entry in docs/dev/UPSTREAM.md cites the LocalRouter commit and PR that added the change.
• Upstream CI runs the zig suites on the three TensorFold PRs. drama-tensorfold has no CI; its PRs list the tests run.

Merge Danger

Door: two-way. Docs only.

Blast Radius: docs. No code or pins change. Our patches stay until each upstream PR merges and our pins move past it; UPSTREAM.md says what to drop when.

🤖 Generated with Claude Code