[hardy30894]

Stacked on #12.

What

secondwind eval --query replays each tool result with the user's latest message as the query, the way the proxy calls the optimizer on live traffic.

What it found

1. The JSON relevance split is value-complete but my check missed it. It keeps the rows that match the query inline in the stub and stores only the rest, so resolve alone returns a subset. The faithfulness check now recognizes a split and requires the inline rows plus the stored rows to be exactly the original's rows, in any order. On the author's 1,491 sessions with --query: 0 of 138,448 blocks unfaithful. A test proves the check also catches one row missing or one row too many.
2. A text version of the split does not pay, so it is not in this PR. I built one (keep the most relevant quarter of the lines inline with their line numbers, store the whole block) and replayed the corpus:

replay	net context saved	reopens per session
no query	24.6%	23.7
query, with a text split	8.5%	21.4

The larger stubs cost far more than the few reopens they prevent, so the query description stays as #12 left it: JSON record arrays only.

Tests

a_relevance_split_is_faithful_across_stub_and_store; the existing eval and gate tests run through the new options struct. Clippy clean.