JVNAUTOSCI-2614: Select minimum-adequate capability plans

[witbrock]

What changed

• Treat registered calls and represented workflows as capability plans on one bounded, cost-sensitive frontier.
• Project declared workflow components, step/component counts, semantic effects, durable-runtime cost, and optional represented direct-equivalence metadata.
• Promote actor-visible components of a semantically matched workflow as plausible simpler plans while leaving semantic adequacy with the adaptive model.
• Apply mechanical dominance only to read-only plans with an explicit authoritative direct_equivalent_capability_names declaration; retain dominated workflows for exact-name inspection.
• Distinguish a read-only workflow's semantic effect from its durable instance state effect.
• Record the selected plan/effect/cost profile and actual transport cost in turn telemetry.
• Replace substring matching and generic one-word prose matches with whole-token, corpus-relative routing evidence, and prioritise structurally declared components over weak lexical matches.
• Document the routing-profile extension and its conservative use.

Why

Workflow relevance was scaled into a separate score range (up to 1000) while direct capabilities received small literal scores. That made representedness an accidental priority and prevented the model from comparing a matched workflow with its simpler direct components. Short substrings and generic routing words also admitted unrelated tools.

The new boundary keeps authority unchanged and does not add a workflow classifier or compulsory controller stage. The adaptive model judges material outcome/evidence adequacy; declared cost breaks ties only after adequacy, and hard pruning requires represented equivalence.

Impact

For the reported Can you list PhD students supervised by me? shape, the canonical entity-retrieval workflow's five read components and the workflow occupy the top six catalogue positions. find_relations_with_argument is fourth, while the workflow remains available when its full five-part evidence contract is genuinely needed.

Validation

• 111 passed post-rebase across adaptive-turn, workflow-turn capability, routing-profile extension, and routing metadata projection tests.
• Ruff and git diff --check pass.
• Broader workflow/discovery validation reached 201 passed before stopping the unrelated Mongo-heavy remainder after 13m46s with no failures.
• Faithful replay with the real default catalogue, canonical entity-retrieval seed, and live tool metadata placed all five declared read components plus the workflow in the top six and recorded the workflow as semantic read-only but operationally durable.
• Live Vontology discovery was read-only exercised but the current capability index reported capability_index_wait_timed_out_build_in_progress; this is a separate availability boundary, not treated as selection success.

Jira: JVNAUTOSCI-2614

[Copilot]

Pull request overview

This PR updates Von’s ordinary-turn capability discovery and ranking so that registered direct tools and represented workflows are treated as comparable “capability plans” on a single, bounded, cost-sensitive frontier. It adds explicit plan/effect/cost metadata to catalogue entries, promotes workflow-declared component tools as plausible simpler alternatives, and introduces conservative mechanical pruning only when a workflow declares an authoritative read-only direct equivalent.

Changes:

• Add workflow plan projection (declared components, step counts, direct-equivalence metadata) and effect profiling (semantic vs operational durable-state effect) to represented workflow capabilities.
• Replace previous substring/generic-token routing with whole-token, corpus-relative evidence and interleaved frontier ordering that keeps both direct and workflow options visible while leaving adequacy to the adaptive model.
• Record selection policy and per-invocation plan/effect/cost context in telemetry traces, and expand/adjust test coverage accordingly.

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
tests/backend/test_workflow_turn_capability_service.py	Adds/updates assertions for workflow plan metadata, semantic vs operational effects, and write/unknown effect marking behaviour.
tests/backend/test_adaptive_turn_service.py	Adds extensive tests for the new frontier ranking, component promotion, rarity-based routing evidence, dominance pruning rules, and selection trace enrichment.
src/backend/services/workflow_turn_capability_service.py	Projects workflow plan/effect/cost metadata into catalogue entries and allows discovery to use registered capability categories to classify semantic effect when evidence is sufficient.
src/backend/services/adaptive_turn_service.py	Implements the cost-sensitive frontier ranking, corpus-relative routing evidence, conservative dominance pruning, compact projection of selection profiles, and selection telemetry enrichment.
docs/engineering/von_workflow_language_manual.md	Documents direct_equivalent_capability_names and the new capability-plan selection/dominance rules.