research: add System One code localization experiment

[BestNathan]

Summary

Adds a second System One research prototype under the existing progressive-action-space topic.

• add a Python code-localization demo: repository -> directories -> files -> source lines -> grounded snippets
• use independent Noul relevance judgments instead of Choice because multiple candidates may be relevant
• preserve full candidate/decision provenance in JSONL traces
• add deterministic fixture tests and a small websocket-oriented fixture repository
• extend the System One topic with the shared StateSpaceGenerator -> CandidateSet -> DecisionPrimitive -> TransitionPolicy -> EvidenceRecorder model
• add a GitHub Actions workflow that runs deterministic demos automatically and collects real System One evidence on manual dispatch

Research evidence

The workflow uploads 90-day artifacts containing:

• run-manifest.json
• Kubernetes result + human-readable execution log
• code-locator result.json
• code-locator trace.jsonl with candidate disclosure, model requests, scores, thresholds, latency, and usage
• stdout.log
• summary.md for real System One runs

The real-model job is manual, uses the existing ds environment, requires TYPESAFE_API_KEY, and accepts subject repository/ref, query, thresholds, and batch size as experiment inputs.

Validation

• existing Kubernetes demo tests are executed by the new workflow
• code-locator tests cover deterministic progressive localization and Noul request construction
• the workflow's fixture job does not require external model credentials

[BestNathan]

Research lineage update: this merged PR remains the historical initial Code Locator prototype in Narness. The later exploration line has moved to BestNathan/system-one-code-explore. Narness convergence and the distilled architectural conclusions are tracked in #26; the in-repo research/code-locator/ tree is now treated as an archived snapshot.