feat: Eval benchmark repo sync to remote targets

[christso]

Objective

Let AgentV sync eval benchmark repos on demand via Studio UI, following an ArgoCD-style model where git is the source of truth and sync is an explicit action.

Current Problem

Today, getting a benchmark repo up to date is left to the caller. Users script ad-hoc clones/pulls per environment. There's no unified way to say "this benchmark lives in a git repo" and have Studio manage it.

Design

1. Add source to benchmark entries

Extend the existing benchmark registry at ~/.agentv/benchmarks.yaml with an optional source field. Uses the existing interpolateEnv() from packages/core/src/evaluation/interpolation.ts.

benchmarks:
  - id: eval-benchmarks
    name: Eval Benchmarks
    path: evals
    source:
      url: ${{ BENCHMARK_REPO_URL }}
      ref: ${{ BENCHMARK_REPO_REF:-main }}
    added_at: "2026-03-20T10:00:00Z"

• source is optional. If absent, path is used as-is (current behaviour).
• No sync field. If source exists, the benchmark is git-backed and syncable.

2. Sync as explicit action (ArgoCD model)

• Studio UI: Benchmarks screen shows a "Sync" button for git-backed benchmarks. Click → oneshot git clone --depth 1 or git pull --ff-only.
• CLI: agentv benchmark sync <id> triggers the same oneshot pull.
• Docker/CI: Run agentv benchmark sync as a pre-step in the container entrypoint or CI pipeline.

No background daemon. No continuous mode. No git-sync dependency.

3. Behaviour

State	Action
No source	Local benchmark. No sync button. Path used as-is.
Has source, first time	git clone --depth 1 --filter=blob:none to path.
Has source, already cloned	git pull --ff-only from source.ref.
Docker/CI	agentv benchmark sync in entrypoint or script.

4. Interaction with existing eval workspace.repos

Individual .eval.yaml files can declare additional repos via workspace.repos. This is per-eval, inline, and unrelated to the benchmark registry. The two coexist:

• Benchmark source — project-level "where does this benchmark live." Drives sync.
• Eval workspace.repos — per-eval "this test needs this additional repo." Used at eval runtime.

Acceptance Criteria

• A benchmark entry with source.url + source.ref can be synced to path via Studio UI "Sync" button or agentv benchmark sync.
• agentv benchmark sync does git clone --depth 1 (first time) or git pull --ff-only (subsequent).
• Existing benchmark entries without source continue to work unchanged.
• ${{ ENV_VAR }} interpolation works in source.url and source.ref.
• Studio benchmarks screen shows sync button for git-backed benchmarks.

Non-Goals

• Continuous sync / background daemon.
• git-sync dependency or binary distribution.
• Two-way sync, conflict resolution, or write-back to source.
• Auto-sync on interval.

[christso]

Closing as superseded.

PR #1234 implemented this design and was closed without merge on 2026-05-13. The follow-up commit ec5168a (refactor: strip sync field and continuous mode from benchmark sync) deliberately removed the sync field, continuous mode, agentv install command, and the git-sync dependency. The parts that were kept ship on current main:

• BenchmarkEntry.source: { url, ref }
• syncBenchmark() — shallow clone on first run, git pull --ff-only on subsequent runs
• agentv doctor command (extensible DEPS registry, currently empty)

The delta this issue proposed (multi-mode sync: field, continuous daemon, agentv install git-sync, bundled git-sync dep) is exactly what was stripped. Per design principles #1 (Lightweight Core) and #5 (YAGNI), re-adding it would walk back a deliberate scope reduction.

If a narrower gap emerges later (e.g., a specific dep that agentv doctor should track, or a real need for continuous sync from a concrete user), file a focused issue against the actual gap.