feat(stage_router): corroborative tanh-scored routing with windowed self-clearing signals

[sabhatinas]

What this PR does

Reworks the stage_router scorer, picker, and signals into a cleaner, honestly-documented router — and moves the routing logic into one shared Rust core.

• One shared Rust core. The scorer + picker now live in switchyard-components/stage_router.rs (a short priority ladder: escalate → de-escalate → scorer → fall-open). Python calls it via a binding (stage_pick_tier); the Rust profile calls it directly. One source of truth. The dead prompt-classification path is deleted. (SWITCH-1056)
• Two-axis scorer. Error signals (a hard error, spinning, or exploring) push toward the strong model; making progress (writing code) pushes toward the weak model. Squashed with tanh, and confidence = |score|.
• Corroborative. One signal alone rarely flips the decision — it takes a signal plus backup. A critical error overrides everything.
• Cleaner signals. spinning and exploring are mutually exclusive. Dropped repeated_cmd_ratio from scoring — it fired on nearly every shell command and never caught a real loop.
• Compaction holds strong. Once the context is compacted, the router pins the strong model.
• Handoff notes. Optional short note at a tier switch telling the next model why it was switched (off by default).
• Per-tier system prompts. Optional prompt injected for the weak or strong tier.

Docs and tests updated; the old calibration scripts are replaced by the scorer skill.

Signals live in libsy. Tool-signal extraction (tool_signals) now lives in switchyard_libsy::ToolSignals; switchyard-components re-exports it — no divergence. The scorer/picker core stays in switchyard-components/stage_router.rs.
count_tokens was split out to its own PR (SWITCH-1048).

Linear: SWITCH-1056 (Rust consolidation), SWITCH-939 (scorer), SWITCH-983 (compaction), SWITCH-982 (handoff notes).

[coderabbitai]

Walkthrough

The stage router now separates tests_passed decisions, uses recent-window progress gates, routes settled runs directly to EFFICIENT, preserves critical overrides, and computes tanh-squashed scores with updated tests.

Changes

Stage router behavior

Layer / File(s)	Summary
Parameterized tanh scoring switchyard/lib/processors/stage_router/scorer.py, tests/test_stage_router_scorer.py	Grouped signal weights and tanh scoring replace fixed weights and hard clipping; tests cover direction, saturation, confidence, and raw contributions.
Routing gates and settled-run precedence switchyard/lib/processors/stage_router/decision_log.py, switchyard/lib/processors/stage_router/dimensions.py, switchyard/lib/processors/stage_router/picker.py, tests/test_stage_router_pickers.py	Recent-window gates and the dedicated tests_passed source support early EFFICIENT routing, while critical severity continues to force CAPABLE; picker scenarios are updated accordingly.

Estimated code review effort: 3 (Moderate) | ~25 minutes

Poem

I’m a bunny with tests in my ear,
Hopping through scores that grow smooth and clear.
Settled runs dash to EFFICIENT light,
Critical thumps keep CAPABLE in sight.
Tanh curls the signals, soft as spring—
New paths bloom while green checks sing.

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Docstring Coverage	✅ Passed	Docstring coverage is 88.24% which is sufficient. The required threshold is 80.00%.
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 captures the main stage_router changes: tanh scoring, corroborative routing, and windowed self-clearing signals.

---

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

[github-actions]

PR Preview Action v1.8.1
🚀 View preview at https://NVIDIA-NeMo.github.io/Switchyard/pr-preview/pr-68/
Built to branch gh-pages at 2026-07-24 18:23 UTC. Preview will be ready when the GitHub Pages deployment is complete.

[ryan-lempka]

@sabhatinas is the plan to refactor this to align with libsy design at a later date / does this take libsy into account?

[sabhatinas]

@sabhatinas is the plan to refactor this to align with libsy design at a later date / does this take libsy into account?

Yes, I plan to migrate to libsy once I have this checked in. I want to have a clean submission of the working algorithm to reproduce results from our submissions.