[TS Calls] Build a real-project call census and model roadmap

Start with the existing native frontend, runner, and observer, then review the inventory in #368 and the related work in #360, #364, and #385. Define the census profile, deterministic aggregation, machine-readable observations, and explicit limits described here. Done means a reviewed P0/P1/P2 model subset is ready for #385, with implementation split into bounded follow-up issues.

Part of #360. Uses #364; contributes to #385.

Goal

Measure unknown calls reached on real TypeScript projects and select a small evidence-based model roadmap.

Scope

• Freeze development-project revisions, licenses, selection criteria, analyzed methods, input domains, frontend/tool versions, seeds and budgets.
• Reuse the existing native frontend, runner and observer. Aggregate decision events outside the engine; #373 snapshots are not required for a basic census.
• Define the observation profile explicitly. A stop-on-first-unknown run cannot reveal later calls on that path. Use a fixed opaque-continuation census and/or a separately reported stop profile to distinguish encountered calls and stopping points.
• Opaque exploration can include infeasible concrete paths and omit callee effects. Report a census of observed events under the profile, not an exhaustive list of all calls or all runtime blockers in a project.
• Normalize API families/methods without conflating same-named user methods with proven built-ins.
• Count projects, distinct call sites, distinct containing functions and dynamic events with explicit denominators. Rank primarily by project prevalence and distinct sites; report loop event counts separately.
• Classify observations as model candidates, frontend/engine defects, external/user behavior or already supported semantics.
• Use #368's inventory to avoid reimplementing existing models. Record candidate domains, residual cases and the rationale for source code versus a genuine engine intrinsic.
• Produce a small P0/P1/P2 list. Select a finite experimental subset before final evaluation; held-out projects never select or tune it.

Definition of Done

• Machine-readable observations preserve profile, limits, project/source examples and raw counts.
• Source-site identity and aggregation are deterministic; repeated budget-limited runs report observed variability instead of promising identical explored events.
• Stop-profile censoring and opaque-profile approximation are explicit.
• A reviewed model subset is ready for #385; missing implementation work is split into bounded family issues with evaluation dependencies.
• No catalog expansion is implemented in this issue.

The integration examples in #365/#366 are allowed before this roadmap; they validate the mechanism and must not be retroactively presented as census-selected models. Corpus selection and aggregation work can proceed independently of #383.