Add focused headless hypothesis runs

[vaalessi]

What changed

• adds focus_thing_id to the official core run, filtering candidates before the global cap and recording the focus in provenance
• adds hypgen-run, a stable file-in/file-out command that returns the canonical hypothesis document, existing cards payload, and a complete current rNPV module request
• requires the valuation frame, comparables, request id, simulation count, and seed rather than selecting or inventing them
• adds explicit LIVE and REPLAY modes with no fallback between them
• vendors and hash-pins the shared interpretability and IndicationThesis contracts

Functional interface

hypgen-run --mode live --input request.json --output result.json
hypgen-run --mode replay --input request.json --output result.json

LIVE always constructs the provider-backed Judge. REPLAY never constructs it and is labeled DETERMINISTIC_REPLAY; it runs the deterministic local pipeline over the caller-supplied graph, not a saved hypothesis fixture.

The result always carries status, execution_mode, output_origin, hypothesis, cards, roi_request, and error. If only ROI adaptation cannot complete, the canonical hypothesis and cards remain available with an explicit reason.

Why

The existing orchestrator imports internal generator state, takes several candidates from one run, and projects them across unrelated contexts. It also builds ROI input by overlaying a fixture. This PR gives the orchestrator one supported, focus-bound call per branch and a complete downstream request.

Checks

• Python 3.13: 336 passed, 3 skipped
• Python 3.12 with current rnpv-roi-calculator installed: 338 passed, 1 skipped
• exact generated roi_request validates as the current ROI ModuleRunRequest
• shared-contract byte hashes and interpretability conformance verified
• compileall passes

No live model call or deployment change was made before the demo.