[TS Calls] Add reproducible unknown-call census

[CaelmBleidd]

The TypeScript engine's unknown-call decisions were not available as a reproducible project-level inventory. This adds runUnknownCallCensus with census and summarize commands: it checks pinned project revisions, loads the complete project scene, records repeated events and explicit complete/partial/timeout/error outcomes, and summarizes unique call sites and containing functions.

The development sample no longer depends on lexicographic file order or BFS. A class qualifies when it has at least minMethodsPerClass ordinary methods with at least minStatementsPerMethod IR statements. Eligible classes and methods receive reproducible SHA-256 ranks from the manifest seed, and methods are taken round-robin across classes. Execution uses CLOSEST_TO_UNCOVERED_RANDOM, records the seed and strategy, and does not stop at entry-method coverage. Synthetic file methods use JacoDB's DEFAULT_ARK_METHOD_NAME.

The manifest uses a 30-second method budget and a 900-second project budget, with up to 40 classes and 40 methods per project. Its three projects are an explicitly limited development pilot. Additional development projects must be pinned before inspecting their results, while held-out projects remain separate from model selection.

The accepted EMPTY_FRESH run at c1e07845 analyzed 97 methods: 45 completed, 36 partial, 16 timed out, and none failed at the tool boundary. The project distribution was 17/40/40 methods across 11/37/14 source files. It recorded 7,002 repeated events at 36 stable source sites in 26 containing functions. Site counts, not repeated exploration events, drive the documented decision; no new model family is admitted by this pilot.

The runner is split into orchestration, seeded method selection, bounded process handling, and event recording. The main runner is 574 lines instead of 1,136.

Validation on the final head:

• JDK 21 compilation and 10 focused census tests passed.
• :usvm-ts:detektMain, :usvm-ts:detektTest, and git diff --check passed.
• Standalone summary regeneration is byte-identical to the generated summary.
• Raw SHA-256: a091dff2d393e131a83e76cc71af51dd38dcd821e2e01ee007cfe8aa56ab6c08.
• Summary SHA-256: 7c59fbfa09af54de1e74ed65f4c762629013688d109305a970b421ee32781d00.
• Manifest SHA-256: 4adba5861339782e1f514f01525561a67aa9f01a6abe914b3d371ff77e1c1bbb.

Optional models are disabled for the census; mandatory semantics and existing pre-dispatch approximations remain unchanged. The 36 partial analyses, 16 timeouts, and three-project scope remain explicit limits on prevalence claims.

Related to #367.