refactor: consolidate input_segments into input (TestMessage[])

[christso]

Objective

Remove the input_segments field from EvalTest by resolving file content directly into input (TestMessage[]) at parse time. Today input and input_segments carry overlapping data — input_segments is a flattened, role-stripped version of input with resolved file text. This duplication caused a bug where {{ input }} in judge templates showed file paths without content (fixed in PR #507 via a render-time patch). Consolidating eliminates the duplication and the patch.

Architecture Boundary

core-runtime — touches the EvalTest type contract, YAML/JSONL parsers, prompt-builder, and evaluator variable binding. All changes are within packages/core/src/evaluation/ and apps/cli/src/commands/trace/.

Current State

Concern	Where it reads from today
Judge template `{{ input }}`	evalCase.input + render-time resolveInputWithFileContent() patch (llm-judge.ts)
Multi-turn prompt building	evalCase.input for messages, evalCase.input_segments for file content lookup
Single-turn prompt building	evalCase.input_segments directly (prompt-builder.ts:165)
Trace scoring file map	evalCase.input_segments for resolvedPath (trace/score.ts)

Design Latitude

• Must: enrich input message content segments with text (file content) and resolvedPath during processMessages() in yaml-parser/jsonl-parser
• Must: update prompt-builder single-turn path to derive segments from input instead of input_segments
• Must: update trace scoring to extract file paths from input
• Must: remove input_segments from EvalTest interface, all parsers, and test fixtures (~18 files)
• Must: remove resolveInputWithFileContent() from llm-judge.ts (the patch becomes unnecessary)
• Free to decide: how to structure the flattening helper for single-turn prompt-builder (inline loop vs extracted utility)
• Free to decide: whether resolvedPath lives on the content segment or is derived at point of use

Acceptance Signals

• input_segments does not appear anywhere in packages/core/src/ or apps/cli/src/
• resolveInputWithFileContent is removed from llm-judge.ts
• bun run build succeeds
• bun test packages/core/test/evaluation/ — all tests pass
• Multi-turn e2e: agentv eval examples/features/multi-turn-conversation/evals/dataset.eval.yaml --target default — 2/2 pass with per-turn scoring and role-annotated context
• Judge templates with {{ input }} show resolved file content (not just paths)

Non-Goals

• Changing the YAML/JSONL user-facing schema (input format stays the same)
• Modifying the TestMessage type definition (content segments are already JsonObject)
• Changing how guidelines are extracted/filtered (that stays in processMessages)

Related

• PR feat: add multi-turn conversation eval example with details field #507 (feature/multi-turn-conversation-eval) — introduced the render-time patch this issue removes