fix(libsy): preserve classifier system instruction

[bbednarski9]

What

Make the libsy LLM capability classifier place its governing policy in
LlmRequest.instructions as a system InstructionBlock. Keep only the task
conversation in LlmRequest.messages, and extend the existing request-builder
test to lock in that separation.

This is a focused libsy correctness fix. It is independent of raw-stream
preservation PR #192 and does not change streaming types, translation APIs, or
Relay code.

Why

switchyard-protocol deliberately separates model instructions from ordinary
conversation messages. The classifier labeled its policy Role::System, but
inserted it into messages. On the OpenAI Chat translation path, ordinary
messages with that role are encoded as user content, so a provider received:

user: <classifier policy>
user: <task to classify>

instead of:

system: <classifier policy>
user: <task to classify>

That breaks the classifier router specifically because its judge call depends
on a library-authored policy prompt. The random router does not synthesize a
judge prompt and is unaffected.

The issue was reproduced through NeMo Relay's real Algorithm::run_stream
dispatcher against the unmodified Switchyard baseline: a provider-boundary
assertion expected the classifier policy to have role system but observed
user. Applying only this change made the same buffered and streaming process
tests pass. A real Inference Hub smoke test then completed classifier and
selected-target calls for both buffered and streaming requests.

This restores the protocol and translation invariant at the source rather than
adding a Relay-specific translation workaround. It is not presented as a
general prompt-injection mitigation.

No linked issue.

How tested

• uv run ruff check . clean
• uv run mypy switchyard clean
• uv run pytest tests/ green (1,723 passed, 35 skipped)
• Manual smoke: NeMo Relay process E2E for buffered and streaming classifier routes, followed by real Inference Hub buffered and streaming routes

Additional Rust validation:

• cargo fmt --all --check
• cargo clippy --workspace --all-targets -- -D warnings
• cargo test -p switchyard-libsy (195 unit, 3 observability, and 4 prompt tests passed)

Checklist

• N/A — this is a Rust-only change to an existing type.
• N/A — no new Python public symbols.
• Unit tests added for new components / bug fixes.
• N/A — no customer-facing surface changed.
• Commits signed off (Signed-off-by: Your Name <email>) per the DCO.

Notes for reviewers

The key invariant is that classifier-owned policy belongs in
LlmRequest.instructions, while the conversation being classified remains in
LlmRequest.messages. Please pay particular attention to preserving that split
as prompt-window behavior evolves.

This branch is based directly on current main and contains one commit changing
only crates/libsy/src/algorithms/llm_class.rs; it is not stacked on PR #192.

Summary by CodeRabbit

• Bug Fixes
  • Improved capability evaluation reliability by separating system guidance from conversation content.
  • Preserved relevant message history while reducing unnecessary context passed during evaluation.
  • Updated validation coverage to ensure requests maintain the expected structure and behavior.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: d34f9f64-e975-4d4f-bf41-bcf41dd0bf88

📥 Commits

Reviewing files that changed from the base of the PR and between cb30e92 and 4c4b885.

📒 Files selected for processing (1)

• crates/libsy/src/algorithms/llm_class.rs

---

Walkthrough

The capability judge request now stores the system prompt in LlmRequest.instructions through InstructionBlock, keeps message history separate, and updates the structured-request test accordingly.

Changes

Capability Judge Request

Layer / File(s)	Summary
Structured request construction crates/libsy/src/algorithms/llm_class.rs	CapabilityJudge::build_request creates a dedicated system instruction and passes the selected messages separately.
Structured request test updates crates/libsy/src/algorithms/llm_class.rs	The test verifies the instruction contents and reduced message count.

Estimated code review effort: 2 (Simple) | ~10 minutes

Poem

I’m a rabbit with a prompt tucked tight,
In instructions, neat and bright.
Messages hop in a separate row,
Tests count the paths they show.
Structured paws applaud tonight!

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly matches the main change: preserving the classifier's system instruction in libsy.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

---

_{Comment @coderabbitai help to get the list of available commands.}

[ryan-lempka]

This makes sense the classifier prompt should be a proper system message. This is how the classifier worked pre-libsy. Nice catch @bbednarski9!