[Stephen Belanger (Qard)]

Evaluation requests now own their output adapter. A public fallible EvalReporter separates terminal, JSONL and CLI SSE presentation from case execution and recording. Local/custom reporters require neither serialization nor Sync; JSONL adds serialization only when selected.

The first reporter error disables that reporter for the run while computation and recording continue. Completion returns a typed reporting stage/index and preserves the computed summary. Source and reporting errors survive together, dropped observers do not cancel reporting, and output failures do not poison client shutdown. summary_to_value now returns serialization errors instead of a null placeholder.

BtEvalRunner selects reporters and submits through the canonical local/recorded entry point. Both local environment flags avoid creating a client, list/filter modes skip execution, and JSONL/SSE replace terminal progress. SSE preserves the existing start/summary/error event format and reports write/flush failures.

Stacked on #147 (refactor/eval-dataset-recording), completing the reporting portion of the original evaluation stage. Remaining CLI completion/lifecycle consolidation belongs to release preparation. Planning notes remain local.

Validation:

• 419 unit/integration tests and 14 doctests passed (4 existing ignored doctests).
• Process-isolated CLI environment/stdout tests cover both local flags, list/filter behavior and clean JSONL. Reporter tests cover errors at every stage, serialization/write/flush failures, independent nonserializable local values, dropped observers and recorded-write/shutdown completion after output failure.
• Strict all-target/all-feature Clippy, rustdoc, formatting and diff checks passed.
• Rust 1.86: 81 focused eval tests, both eval examples and 6 eval doctests passed using an isolated compatible dependency resolution. The default dependency graph's existing MSRV mismatch remains release-preparation work.