fix(libsy): validate classifier scores, fix packaging and strict rustdoc

[elyasmnvidian]

Three fixes to the libsy LLM-classifier router (crates/libsy/src/algorithms/llm_class.rs), plus a packaging and a rustdoc fix.

Streaming classifier verdicts were ignored. The router read the score with as_agg(), which returns None for a streaming response, so a classifier target that streamed its score always fell open to the strong model no matter what it returned. It now drains the stream with into_agg() and routes on the real score. A genuine stream error still propagates instead of being swallowed.

Invalid scores were trusted as verdicts. A classifier reply that parsed as a float but was not a probability — NaN, ±inf, or anything outside [0, 1] — was compared against the threshold directly. NaN and negative values compared below the threshold and were routed to the weak model. The router now keeps only scores in [0.0, 1.0]; everything else fails open to the strong model, same as an unparseable reply.

Packaging: cargo package -p switchyard-libsy now succeeds. The switchyard-protocol path dependency carries version = "0.1.0", which cargo requires before it will package a crate that has a path dependency.

Docs: RUSTDOCFLAGS="-D warnings" cargo doc is clean again. Fixed two intra-doc links that pointed at private items (DriverRequest, Driver::stream) and dropped redundant explicit link targets in lib.rs.

New regression tests cover the streaming drain and the invalid-score fail-open.

Note for reviewers: LlmClassifier is a Rust library type. The Python bindings (crates/switchyard-py/src/libsy_bindings.rs) expose only noop and random, so this router is not reachable through switchyard serve yet — the tests are the proof of behavior.

How to try it: cargo test -p switchyard-libsy runs the new tests.

[coderabbitai]

Walkthrough

Changes

LLM classifier routing

Layer / File(s)	Summary
Classifier scoring and routing tests crates/libsy/src/algorithms/llm_class.rs	Classifier responses are fully aggregated before scoring, probabilities are restricted to [0.0, 1.0], and buffered, streaming, invalid-score, and threshold routing cases are covered.
Crate metadata and documentation crates/libsy/Cargo.toml, crates/libsy/src/core/algorithm.rs, crates/libsy/src/lib.rs	The protocol dependency adds version metadata, while documentation wording and intra-documentation links are updated.

Estimated code review effort: 3 (Moderate) | ~20 minutes

Poem

I’m a rabbit with a streaming score,
Draining every delta to the floor.
NaN hops out, the bounds stand clear,
Strong or weak now routes sincere.
Tests twitch whiskers: all is right!

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
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.
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title accurately summarizes the main changes: classifier score validation, packaging fixes, and rustdoc cleanup.

---

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

[coderabbitai]

Actionable comments posted: 1

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@crates/libsy/src/algorithms/llm_class.rs`:
- Around line 417-419: Remove the tracker reference “NvBug 6485976” from the NaN
fail-open comment at crates/libsy/src/algorithms/llm_class.rs lines 417-419
while preserving its rationale; likewise remove “NvBug 6485975” from the
streamed-score rationale at lines 439-441 without changing the surrounding
behavior or explanation.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 81f87868-d92e-4236-97ce-2e768d7abaac

📥 Commits

Reviewing files that changed from the base of the PR and between 2fa6b87 and 2e417fe.

📒 Files selected for processing (4)

• crates/libsy/Cargo.toml
• crates/libsy/src/algorithms/llm_class.rs
• crates/libsy/src/core/algorithm.rs
• crates/libsy/src/lib.rs

[elyasmnvidian]

Before/after — this fix is library-mode, so the proof is the classifier's own tests

LlmClassifier lives in crates/libsy and is not reachable through switchyard serve (the Python bindings expose only noop and random). The triggers — an invalid score, or a streamed verdict — also can't be produced on demand by a real gateway model. So the reproduction is the regression test: red on main's logic, green on the fix.

I ran the branch's llm_class tests twice: once after reverting the score extraction to main's version (as_agg() + no range filter), once with the fix.

With main's logic (as_agg(), no [0,1] filter):

out_of_range_score_defaults_to_strong ......... FAILED   (NaN routed weak)
below_range_score_defaults_to_strong .......... FAILED   (-0.1 routed weak)
streamed_score_below_threshold_routes_weak .... FAILED   (streamed "0.2" dropped -> strong)
test result: FAILED. 4 passed; 3 failed

With the fix (into_agg() drains the stream; .filter(|s| (0.0..=1.0).contains(s))):

test result: ok. 7 passed; 0 failed

Command: cargo test -p switchyard-libsy --lib llm_class. cargo clippy -p switchyard-libsy --all-targets is clean.

Audit follow-up applied in this push: added below_range_score_defaults_to_strong (a parseable -0.1 that main routes weak).