VALUE-04A: distinguish decision-grade device evidence

[daniele21]

Parent: #145

Outcome

Prevent contextual telemetry from being promoted to model-resource evidence merely because it has HOST/RUNTIME provenance.

This isolated VALUE-04A slice introduces the canonical application classification boundary for decision evidence:

• black-box CLIENT performance remains attributable to the frozen execution and can participate after runtime comparability;
• current host-stdlib-v1 telemetry remains context-only because it measures the Performance Lab process/host rather than attributable model-server resource consumption;
• current local-llm-server-status-v1 activity/concurrency telemetry remains context-only;
• HOST/RUNTIME resource metrics require explicit opt-in from an owning versioned telemetry contract before they can establish the Resources dimension as evidence-available;
• zero completed candidate Runs no longer become evidence-available through all([]) semantics.

Campaign compatibility now exposes an additive evidence_note explaining why a dimension is unavailable. The TypeScript boundary accepts the additive field without forcing existing mocked payloads to manufacture it.

Files

• src/performance_lab/application/device_evidence.py
• src/performance_lab/application/campaign_queries.py
• src/performance_lab/application/campaign_models.py
• frontend/src/api/campaign-types.ts
• tests/test_device_evidence.py
• tests/test_campaign_queries.py

Invariants

• Quality / Performance / Resources remain separate.
• Unknown/unattributable resource evidence remains unavailable, never zero.
• Hardware/runtime fingerprint compatibility is still checked before any future resource comparison.
• The current strict-quality-dominance recommendation policy is unchanged; VALUE-04C will version policy semantics separately.
• No guessed sensor or metric contract is introduced.

Dependency / publication gate

Prepared from dev@0b9c376f2dbcbb167f8df8488374c1b0c5d0ace2 under the adopted repo-template-sw 0.9.1 baseline. Keep draft and do not integrate until VALUE-02D #132 PASS / VALUE-02 DONE.

Evidence boundary

Iteration CI proves deterministic classification/projection semantics only. Representative resource claims remain VALUE-04D / REAL_ENVIRONMENT evidence.