fix: resolve routing tier for fallback and affinity-reuse decisions

[thossullivan]

What

Resolve the routing tier for fallback and affinity-reuse decisions in llm_classifier cascades, so switchyard.requests (and the routing log) carry a tier attribute on every decision path. This matches what the context-overflow fallback already does.

Closes #276

Why

When the judge abstains or is unavailable, the cascade's DefaultTarget decides; on sticky turns, AffinityRouter decides. Neither implements Classifier::routing_tier, so the trait default returns None (crates/libsy/src/core/classifier.rs:75-77) and the metric is emitted with no tier attribute (crates/libsy/src/algorithms/fall_through.rs:294, crates/libsy/src/observability.rs:548-551). The same strong model then lands in two time series, one labeled and one not, and consumers cannot unify them. We hit this monitoring a governed dispatch system: one agentic session showed 1 request labeled tier="weak" and 45 on the strong model with no tier at all (mechanism: #272).

The overflow fallback already resolves the tier across the cascade via find_map (fall_through.rs:221-233). This PR applies the same pattern to the primary decision path. Because TaskClassifier::routing_tier maps model name to tier, both halves get the tier of the model actually selected: a judge-abstain fallback gets tier="strong" and an affinity reuse gets the pinned model's tier. No new tier values, no trait changes.

Related: #205 (in flight) answers how often the classifier failed open and why, with a dedicated counter; this PR answers which tier actually served. They compose, and we are glad to rebase over #205 whenever it lands (same test file). Prior art considered: #31, #135, #168.

How tested

• uv run ruff check . clean (no Python changes)
• uv run mypy switchyard clean (no Python changes)
• uv run pytest tests/ green: 1330 passed, 31 skipped (no Python changes)
• Manual smoke: patched release switchyard-server, classifier route with a weak and strong pair over OpenRouter. A trivial request classified weak (switchyard_requests_total{model="deepseek/deepseek-v4-flash-0731",tier="weak"} 1). A tool-continuation request, whose newest message is a tool result the judge cannot classify, fell back (switchyard_requests_total{model="openai/gpt-5.6-sol",tier="strong"} 1). Before this change the second series had no tier attribute.
• Rust on the pinned toolchain (1.96.1): cargo fmt --check and clippy clean; cargo test -p switchyard-libsy green (231 tests). The two new tests, judge_unavailable_fallback_keeps_the_tier_label and affinity_reuse_keeps_the_tier_label, fail without the fall_through.rs change.

Checklist

• One class per file; filename = snake_case of the primary class. (n/a, Rust only)
• New public symbols exported from switchyard/__init__.py.__all__ if intended for downstream use. (n/a)
• Unit tests added for new components / bug fixes.
• README / --help updated if customer-facing surface changed. (n/a, this completes an existing metric attribute; glad to add a line to the metrics reference doc if wanted)
• Commits signed off (Signed-off-by:) per the DCO.

Notes for reviewers

• The or_else runs only when the deciding classifier reports no tier, so classified decisions are unchanged, and the change is inert for routes without tiering (every classifier returns None and the field stays None).
• Expect a tests/observability.rs conflict with feat(libsy): count classifier fail-open fallbacks on /metrics #205; glad to rebase in either order.
• The 45-unlabeled-requests session traces to the judge receiving tool-continuation turns. That mechanism is [bug] llm_classifier: agentic tool-continuation turns silently fall through to the capable target because the judge is handed the newest tool result #272, with a self-contained reproduction attached; this PR only makes the fallback visible and labeled.

Summary by CodeRabbit

• Bug Fixes

  • Improved request routing when the primary classifier does not provide a routing tier.
  • The system now checks fallback options and selects the appropriate tier for the resolved target while preserving the primary decision when available.

• Observability

  • Added coverage to ensure fallback routing and session affinity report accurate tier labels and maintain consistent metrics across repeated requests.

[coderabbitai]

Walkthrough

FallThrough::route now finds a tier from the classifier cascade when the deciding classifier lacks one. Observability tests cover unavailable-judge fallback and session-affinity tier labels.

Changes

Tier Routing and Observability

Layer / File(s)	Summary
Cascade tier resolution crates/libsy/src/algorithms/fall_through.rs	FallThrough::route checks the classifier cascade for a tier associated with the resolved target when the deciding classifier provides none.
Fallback and affinity observability crates/libsy/tests/observability.rs	Added a client stub and tests for classifier timeouts, strong-tier fallback, and weak-tier session-affinity metrics.

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

Poem

A rabbit watched the tiers align,
Through cascading paths they shine.
When judges pause, the route goes on,
Strong and weak labels both stay drawn.
Two requests share one metric trail,
Hop, hop—fallbacks never fail!

🚥 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 clearly summarizes the main change: resolving routing tiers for fallback and affinity-reuse decisions.

---

_{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/tests/observability.rs`:
- Around line 1233-1245: Extend the observability test around the two-request
loop to track non-routed judge calls in ClassifierClient, then assert that the
second request does not increase that count. Keep the existing routing_tier
assertions, and use the recorded call count to verify the second request reuses
the session-affinity decision path rather than invoking the judge again.

🪄 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: d2d5e19e-2405-44c7-8282-6f94390ca9cb

📥 Commits

Reviewing files that changed from the base of the PR and between a9c04b3 and 50c8be2.

📒 Files selected for processing (2)

• crates/libsy/src/algorithms/fall_through.rs
• crates/libsy/tests/observability.rs

[ayushag-nv]

@thossullivan Thanks for putting up the PR. Can you put up a Issue and link a PR to that.

[thossullivan]

Filed #276 for this and linked it in the description. Thanks for the look!