[plotarmordev]

Description and motivation

I move the full-model engine pin from TensorFold d1018b31 to v1.1.9 (465e48b6). The full-model patch series is now empty. v1.1.9 already includes every patch I carried:

Old patches	In TensorFold since	Upstream
0001–0011 (prefill rows, replay fix, frame budget, controls allowlist, voice catalogs, silence guard and redraws)	v1.1.4	#1119, #1120, #1134, #1135, #1152, #1157–#1162
0012 reference-prefill reuse	v1.1.5	a06a4dff, #1237
0013/0014 request sampling controls and JSON boundaries	v1.1.6	#1277 (e19b8073, 90082556)

v1.1.9 also includes the slot-release fix (7e43ebeab, #1438), which shipped in v1.1.8. So this bump needs no local patch for it.

Patch check against v1.1.9: 0013 and 0014 no longer apply with git am, because the files they add (docs/speech-sampling.md, zig/src/server/speech_sampling*.zig) already exist there. I applied 0001–0014 to d1018b31 and compared the result with v1.1.9:

• docs/speech-sampling.md is identical.
• speech_sampling.zig and its test differ only by upstream's later fix 4a556244, which sends a temperature of -0/-1e-400 to the sampler as +0.
• v1.1.9's zig/src/server/root.zig registers both test files.

The same is true for the dialogue-speaker patch 0015 that #28 added: v1.1.9 already ships it (#1278), so this branch removes it.

What changes:

• engine/UPSTREAM now pins v1.1.9. The 14 full-model patches are removed. engine/patches/SHA256SUMS stays as an empty file, which declares the empty series.
• scripts/build-engine.sh accepts an empty series only for --weights full, and only when the manifest is present and empty. An empty NVFP4 manifest, a missing manifest and a manifest that lists a patch that is not there all still fail before checkout. A model-free build test covers these cases.
• recipe/pilot.py passes TF_DRAMA_PREFILL_ROWS=128 for full and custom checkpoints on the Linux native backend. Upstream defaults to single-row prefill, and old patch 0001 turned batched prefill on. NVFP4 and the sglang backend are unchanged.
• README, engine/patches/README.md, CONTRIBUTING, NOTICE, the macOS and sampling docs and CHANGELOG now describe the new pin. They also name its two behaviour differences:
  • With the mid-silence rule on, X-Silence-Mid-Trips now counts only the returned take. This has been true since v1.1.4 and is unchanged through v1.1.9.
  • The -0 temperature sign fix above.

The NVFP4 pins (CUDA and Metal) and their patch series are untouched.

This replaces #31 (same change on the old base) and the earlier 1.1.4 bump branch, neither of which I will merge.

Rebased onto main after #28 (f68738ea):

• [3/3] Preserve dialogue speaker roles across all pinned Drama engines #28's full-model patch 0015-dialogue-speakers.patch, its SHA256SUMS line, its patch-README paragraph and its README row are gone. The row moves into the "In TensorFold 1.1.9" table, citing TensorFold #1278, which landed upstream as 715a074dd, an ancestor of v1.1.9.
• recipe/README-dialogue.md now says the frame-cap stopping result was measured on the previous pin (d1018b31 plus patches 0001–0015) and that v1.1.9 is untested for it, tracked in Full-model speaker-tagged dialogue can fail to emit its stop #32.
• README (speaker tags) and CHANGELOG (Added) point to the known full-model stopping limitation and Full-model speaker-tagged dialogue can fail to emit its stop #32.
• results/dialogue-stopping-20261030.json records the NVFP4 build's source_tree (1ff01f58…, the tree of fork commit e06a0cda) beside its commit. No measured value changes.
• The CUDA and Metal NVFP4 0003/0006 dialogue patches from [3/3] Preserve dialogue speaker roles across all pinned Drama engines #28 stay.

---

Testing

CPU only. I built and tested on a GB10 host, without using its GPU. The engine builds below ran on #31's head cb958306; the rebase adds only docs, the result note and the removal of 0015, so the B build input (465e48b6 + 0 patches) is unchanged.

• Engine builds, ZIG=zig-0.17.0 NVCC=/usr/local/cuda/bin/nvcc JOBS=4 scripts/build-engine.sh --weights full <dir>:
  • A, on main db6ba7da: d1018b31 + 14 patches = e0f55c97. Exit 0 in 3:46. tensorfold-native sha256 93fc4ca1…d249, reports version 1.1.1.
  • B, on this branch cb958306: 465e48b6 + 0 patches = 465e48b6. Exit 0 in 4:32. tensorfold-native sha256 ca64be09…9c89b, drama3-generate sha256 32bbb744…6027, reports version 1.1.9.
  • The B binary's serve --help lists the guard flags the recipe uses. --voices, --allowed-controls and --max-frames don't appear in the help text on A or B; I found them in v1.1.9's zig/src/server/cli.zig.
• python3 -m unittest discover -s recipe -p 'test_*.py' on this head: 50 tests OK, 15 skipped (the explicit macOS subprocess smoke tests).
• sha256sum -c SHA256SUMS on this head: CUDA NVFP4 4/4 OK, Metal NVFP4 6/6 OK; the full-model manifest is empty with no patch files.
• The new build test fails against main's scripts/build-engine.sh (1 failure) and passes with this change.
• python3 bench/stream_check.py --self-test on this head: passed.
• bash -n scripts/build-engine.sh: OK.
• python3 gpu-checks/runner.py --patch-selftest and --diag-selftest on this head: PASS (B full runs no checksum step, a changed NVFP4 patch or build script aborts the identity check, a build from another pin is not reused, a prep abort writes an INCONCLUSIVE summary).
• zig build test -Dnvcc=/usr/local/cuda/bin/nvcc in the v1.1.9 engine tree: exit 0 in 9:23. The test output includes expected-failure lines from negative cases. The build host exposed no CUDA device, so this covers host-side tests only.

GPU A/B pins (gpu-checks/runner.py): A = f68738ea (main: d1018b31 + full-model patches 0001–0015 and the CUDA NVFP4 series). B = 3f44de62 (this branch's v1.1.9 tree; a commit cannot name its own SHA. Its engine/, scripts/, recipe/ and bench/ equal this head's; the runner itself runs from the synced gpu-checks/). I compare against main because this pull request claims v1.1.9 matches the current 0001–0015 series; an older A would count 0013–0015 as v1.1.9 differences. Both trees take the same requests: seeds and speaker tags on A too, and the same tagged catalog. Each full build must reach its one deterministic HEAD (A 3e1721df, B 465e48b6); a build from any other pin is never reused. A and B pin the same NVFP4 engine: the runner checks that from the git tree of engine/nvfp4 and the two reviewed build-engine.sh blobs before any build, builds no NVFP4 engine and runs no NVFP4 A/B check. gpu-checks/plan-output.txt and diag-plan-output.txt are regenerated from --plan / --diag --plan.

Not run yet: GPU A/B. I will run gpu-checks/run-all.sh on the device when its GPU is free. On both trees with the full model it runs: G1, the streaming check, two corpus capture passes and the reference-cache transitions, then compare_gate.py on the captures, whose byte-identity true/false is recorded as a result; G2, request sampling (seed repeatability, temperature, top_k, top_p, and a 400 for top_k 65); G3, the tagged and untagged dialogue line with the builtin and the tagged catalog voice. run-diag.sh adds D1, the issue #32 frame-cap count for that dialogue at seeds 1–5 on both trees, and D2, Whisper word error rates for D1's replies. I will add the results here. Until then this pull request has no audio-identity or qualification evidence, and I expect no byte identity with A: upstream changed the engine between d1018b31 and v1.1.9.

---

Performance or quality claim

None.

---

Checklist

• One change per patch, new patches at the end of the series, SHA256SUMS and engine/patches/README.md updated (series emptied; manifest kept empty)
• scripts/build-engine.sh applies the whole series and builds from a clean checkout
• Every behaviour change has a deterministic test that fails without it
• No dead code, debug output or leftovers
• No personal voices, recordings, private names, host names or private paths in code, fixtures, docs, evidence or commit messages
• Where audio should not change, compare_gate.py reports "identical": true and stream_check.py passes; where it changes on purpose, the qualification gates are reported before and after (pending the GPU A/B)

[plotarmordev]

GPU A/B on DGX Spark (CUDA), head 68c6308, A = main f68738ea, B = 3f44de62 (TensorFold v1.1.9, no full-model patches). Both full builds reached their pinned HEADs.

Check	Result
G1 corpus captures, byte comparison A vs B	PASS, byte-identical
G2 request sampling (seeds, temperature 0, top-k 1)	PASS
G3 tagged and untagged dialogue	PASS

Status: complete (wall time about 40 min including both builds). I haven't listened to the audio yet. The issue #32 frame-cap diagnostic (D1/D2) on both trees is running now; I'll post it here when it finishes, and I'm holding the merge until then.

[plotarmordev]

Issue #32 frame-cap diagnostic (D1/D2), same pins, CUDA full model, seeds 1–5:

Tree	Tagged dialogue capped	Untagged capped
A main (previous pin + 0001–0015)	3/5	0/5
B TensorFold v1.1.9	3/5	0/5

Per-reply bytes, durations and RMS are identical between A and B, so v1.1.9 behaves exactly like the previous pin here: the known stopping limit is unchanged, not fixed. D2 (Whisper) completed. Status: complete.

Exact-head local checks on 68c6308: 50 unit tests OK (15 skipped), stream_check self-test, patch and diag self-tests, NVFP4 checksums all pass. No hosted CI runs on this repo. Merging now; I'll follow up with a small docs PR so README-dialogue states the v1.1.9 measurement instead of "untested".