chore(deps): [cherry-pick 1.5.1] bump dynamo-renderer to 5.1.2 and dynamo-protocols to 5.4.1 (#14595, #14154)

[nv-anants]

add #14595 to release

[dmitry-tokarev-nv]

Reviewed as a cherry-pick of #14595 onto release/1.5.0. Everything below was run locally against head f67175cb on macOS ARM64 (cargo test -p dynamo-llm --lib --no-default-features --locked), not read off the diff.

What the cherry-pick actually contains

gh api repos/ai-dynamo/dynamo/compare/release/1.5.0...f67175cb is two upstream commits, not one: #14595 (dynamo-renderer 5.1.0 -> 5.1.2, dynamo-tokenizers 1.8.0 -> 1.8.1) and #14154 (dynamo-protocols 5.4.0 -> 5.4.1).

The second one is required, not stray churn: the crates.io sparse index records dynamo-renderer 5.1.2 as depending on dynamo-protocols ^5.4.1, and this workspace pins that exactly, so 5.1.2 cannot resolve against =5.4.0. On main it was already present (#14154 merged a week before #14595), which is why #14595's own diff is Cargo.toml +2/-2 while this one is +3/-3 — and the extra line is exactly dynamo-protocols. The cherry-pick is faithful to both.

Verified

• The regression test pins the bump. Reverting only Cargo.toml and the three lockfiles to the merge-base (08a2c4f6) while keeping the new test, test_nonleading_system_message_normalized_for_strict_template fails on renderer 5.1.0 with InvalidArgument: invalid operation: System message must be at the beginning. (in tool_use:1), and passes on 5.1.2. Control on the same tree: test_restrictive_template_accepts_request_with_user_message passes both ways, so that is the bump and not the environment.
• The lockfiles are genuine resolutions, not a hand-merge. All three checksums match index.crates.io exactly (dynamo-renderer 5.1.2 -> 3ae2eaa1..., dynamo-tokenizers 1.8.1 -> ed232a9d..., dynamo-protocols 5.4.1 -> dbaf3179...), and cargo metadata --locked --no-deps is accepted for the root workspace and for both binding workspaces (lib/bindings/python, lib/bindings/kvbm).
• One pin site, no drift surface. All three are [workspace.dependencies] entries consumed via workspace = true (lib/llm, lib/backend-common, deploy/inference-gateway/ext-proc); grep finds no second copy of any of the three version literals outside Cargo.toml and the lockfiles, and docs/fern/scripts/tests/test_gen_rust_api.py discovers versions rather than hard-coding them.

Two P2s inline, both about the gap between what the one-line body promises and what this lands on the release branch. Neither asks for a code change. No P0/P1.

I also replied on the existing dynamo-review-agent thread with a measured refutation of its premise.

[dmitry-tokarev-nv]

Approving. Zero P0/P1; the two outstanding items are both P2 and both ask for description text, not a code change.

Verified by execution (macOS ARM64, head f67175cb):

• The cherry-pick is faithful and complete. compare/release/1.5.0...f67175cb carries #14595 and #14154. The second is required, not stray: dynamo-renderer 5.1.2 declares dynamo-protocols ^5.4.1 in the crates.io index against an exact workspace pin, so 5.1.2 cannot resolve on =5.4.0. The delta from #14595's own main diff (Cargo.toml +2/-2 there, +3/-3 here) is exactly that one line.
• The new test pins the bump. With only Cargo.toml and the three lockfiles reverted to the merge-base 08a2c4f6 and the test kept, test_nonleading_system_message_normalized_for_strict_template fails on renderer 5.1.0 (InvalidArgument: invalid operation: System message must be at the beginning. (in tool_use:1)) and passes on 5.1.2. Control test_restrictive_template_accepts_request_with_user_message passes on both trees.
• The lockfiles are real resolutions. All three registry checksums match index.crates.io, and cargo metadata --locked --no-deps is accepted for the root workspace and both binding workspaces.
• One pin site. All three versions live only in [workspace.dependencies] and reach consumers via workspace = true; no second copy of any version literal exists outside Cargo.toml and the lockfiles.

Remaining, both P2, both non-blocking:

1. Cargo.toml:71 — the body says "add #14595 to release" but this also lands #14154's client-visible /v1/chat/completions serialization change (nulls omitted, refusal always present) on release/1.5.0, and no other cherry-pick PR carries it here.
2. Cargo.toml:63 — 5.1.1 rides along and changes the default DeepSeek V4 prompt: measured 73 -> 549 characters for an identical request with no reasoning_effort, inserted at index 0.

Neither needs a code change. Please fold both into the description so they reach the 1.5.0 release notes.

Deliberately deferred: I ran the targeted test plus one control rather than the full dynamo-llm suite locally — rust-tests (.) is green on this head and covers it. The DeepSeek V4 evidence is renderer-level rendering through deepseek_formatter_for, the exact entry point prompt_formatter_from_mdc uses; I did not exercise it end to end against a served V4 model.

I also replied on the dynamo-review-agent thread: its premise does not hold. The normalization is gated on a load-time probe, a template that accepts a non-leading system is left untouched, and where demotion does happen the pre-bump behaviour was a hard 400 — so the change lowers that content's authority rather than raising it.

[dagil-nvidia]

Held for Dynamo 1.5.1. release/1.5.0 is closed at RC11, the 1.5.0 GA build, so this cherry-pick must not merge there. Converted to draft; it retargets to release/1.5.1 as soon as that branch is cut (patch targeted for 2026-09-30). Tracked on the Dynamo 1.5.1 cherry-pick canvas.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.