Baseline agent-recall metrics

[catoncat]

Parent

#86

What to build

Add a baseline report for agent-recall work. The report should measure the current find -> read-* workflow before recall-packet, projection, or dogfood changes land. It should make operation cost, DB size, result count, and returned-context size visible in one reproducible artifact.

Acceptance criteria

• A benchmark/eval command reports search latency, result counts, output byte/char counts, and read-context size for representative recall queries.
• The report includes DB size and table-size breakdown for the active Sherlog DB.
• The report records current dogfood scorecard summary when a dogfood file is supplied.
• The report output is machine-readable and also has a compact human-readable summary.
• Existing perf/eval commands continue to pass.

Blocked by

None - can start immediately

[catoncat]

Implemented and merged in PR #94.