Add *.eval.ts auto-discovery so TS-first eval authors don't need a hand-wired runner

[christso]

Problem

Running evals authored in TypeScript today requires writing a run.ts that imports case modules and calls evaluate(...) explicitly. YAML evals are auto-discovered by the CLI; TS evals aren't.

Evidence: apps/cli/src/commands/eval/commands/run.ts:23 describes the positional args as "Path(s) or glob(s) to evaluation .yaml file(s)". Globbing works for YAML / JSONL / JSON only. The sdk-config-file example covers agentv.config.ts discovery but that's the config, not eval case files.

That missing piece of boilerplate is the main reason the TS authoring path feels second-class next to YAML.

Proposal

A discovery convention:

• CLI discovers **/*.eval.ts (and **/*.eval.js after a build step, if relevant) the same way it discovers EVAL.yaml / *.eval.yaml.
• Each discovered module default-exports (or named-exports) an EvalConfig value, and the CLI runs it with the same reporter, inspect, and compare tooling as YAML evals.
• Config discovery precedence and --filter / --tag / --only flags apply uniformly across YAML and TS evals.
• Runtime: use whatever loader the runtime supports for TS modules (Bun direct import, or tsx / jiti for Node). Document the expectation.

Acceptance criteria

• agentv run picks up *.eval.ts files with no extra flags.
• A TS eval and a YAML eval in the same workspace produce identical trace / inspect output.
• Example under examples/features/ demonstrates a mixed YAML + TS suite.
• Docs updated with discovery rules and the runtime expectation for executing .ts files.
• --workers, --threshold, --tag, --exclude-tag, --filter, --retry-errors, --output, and the cache/output-dir behaviour all work identically for TS-authored suites.

Non-goals

• Does not add new EvalConfig fields — depends on Close programmatic TS API gap: add beforeAll, budgetUsd, and multi-turn fields to EvalConfig / EvalTestInput #1115 for beforeAll / budgetUsd / multi-turn parity first.
• Does not replace the programmatic evaluate() function; .eval.ts files use it under the hood.

Depends on

• Close programmatic TS API gap: add beforeAll, budgetUsd, and multi-turn fields to EvalConfig / EvalTestInput #1115 — without the programmatic API gap closed first, auto-discovery would surface a TS authoring path that's still missing beforeAll / budgetUsd / multi-turn support.

Motivation

Closes the DX gap with hand-rolled TS harnesses while keeping agentv's framework advantages — cost tracking, inspect, compare, grader variety. The current cliff is: "use YAML, or write your own runner script." Neither is what a TS-first user wants.