feature: support hierarchical projection composition and dependency ordering

[Xunzhuo]

Describe the feature

Support hierarchical projection composition so later projection nodes can consume earlier projection scores, calibrated confidences, or named projection outputs instead of forcing every routing overlay into one shallow scores -> mappings step.

Primary layer

global level

Why this layer?

This changes the generic projection contract and runtime evaluation order across router config, DSL, CLI schema, and insights surfaces. It is broader than one detector family or one UI page.

Why do you need this feature?

Today projection composition is shallow:

• scores read base signals
• mappings read one score
• decisions read named mapping outputs

That works for simple coordination, but it gets awkward for layered routing constructs such as:

• difficulty -> verification pressure -> premium reasoning overlay
• ambiguity -> escalation pressure -> fallback policy band
• cost pressure + quality pressure -> final policy label

Without projection-to-projection composition, those interactions either get flattened into one oversized score or pushed back into decision logic.

Additional context

Affected repo surfaces

• src/semantic-router/pkg/classification/classifier_projections.go for dependency-aware runtime evaluation
• src/semantic-router/pkg/config/{projection_config.go,validator_projection.go} for graph-aware validation
• src/semantic-router/pkg/dsl/{ast.go,parser.go,compiler.go,decompiler.go,validator_conflicts.go} for authoring and compile/decompile parity
• src/vllm-sr/cli/{models.py,validator.py} for CLI parity
• dashboard/frontend/src/pages/ConfigPageProjectionsSection.tsx if dashboard editing exposes dependent projection nodes

Suggested starting points

• Start by defining one constrained dependency model, such as score inputs that may reference earlier projection scores or outputs.
• Add validator rules for undefined references, cycles, and evaluation ordering before broadening the syntax.
• Keep the runtime auditable by making ordering explicit rather than implicit.

Current contract limits

The current runtime computes projections in a single shallow pass:

1. partitions normalize competing base-signal matches
2. scores compute weighted sums
3. mappings emit decision-visible labels

That means scores do not currently form a richer typed dependency graph.

Validation commands

make agent-report ENV=cpu CHANGED_FILES="src/semantic-router/pkg/classification/classifier_projections.go,src/semantic-router/pkg/config/projection_config.go,src/semantic-router/pkg/config/validator_projection.go,src/semantic-router/pkg/dsl/ast.go,src/semantic-router/pkg/dsl/parser.go,src/semantic-router/pkg/dsl/compiler.go,src/semantic-router/pkg/dsl/decompiler.go,src/semantic-router/pkg/dsl/validator_conflicts.go,src/vllm-sr/cli/models.py,src/vllm-sr/cli/validator.py,dashboard/frontend/src/pages/ConfigPageProjectionsSection.tsx,dashboard/frontend/src/pages/configPageSupport.ts"
make vllm-sr-test
make test-semantic-router
make dashboard-check

Acceptance criteria

• The projection contract can express at least one projection node consuming an earlier projection artifact.
• Runtime evaluation uses a deterministic dependency order and rejects cycles.
• Config validation, DSL, CLI schema, and decompile/compile flows all support the new dependency model.
• Docs or maintained examples show one layered projection flow end to end.
• Unit tests cover dependency ordering and cycle rejection.

Non-goals

• arbitrary scripting or free-form expression languages in projections
• recursive or cyclic projection graphs
• bypassing decision readability by pushing full routing policy into projections

Related

• Child of feature: track projection-layer evolution #1711.
• Related to feature: allow projections to consume raw numeric signal values #1757, but separate: feature: allow projections to consume raw numeric signal values #1757 expands input value channels, while this issue adds projection-to-projection composition.