v1.53.0 - recall that knows the question

Open Second Brain v1.53.0 - recall that knows the question

Grown from #181 by @Yori940619, whose diagnosis was right: Hermes prefetch recalled the same query-blind context pack on every gated turn, no matter what the user actually asked. The contribution proposed routing prefetch through raw search; the maintainer review moved the mechanism into the curated lane instead, and the redesign ships here with the author's commit as the base of the branch.

brain_context_pack gains a ranked query mode: with query_mode: "ranked" the pack orders its already-collected candidates by deterministic lexical relevance to the query - the same stopword-free kernel the decision matcher uses, no model, no embeddings - within the existing tier order, so a peripheral page still never outranks a core one and nothing is excluded; the token budget decides, and an absent argument stays byte-identical. Because the ranking lives inside the pack lane, the injection guard, the owner scope, the server-side token budget, and the context receipt from v1.52.1 are all inherited rather than bypassed: the Hermes provider simply passes the user's query to the call it already made, and the server receipt remains the outcome sample id.

Along the way the pack projection stopped discarding its own warnings - the one surface that puts vault text in front of a model every turn can now say that an injected memory is contested - and the provider names the lane it recalled through, surfaces pack warnings instead of swallowing them, scopes its outcome-call defaults to posting only, and drops a local character budget in favor of the server-side one, pinned by a Cyrillic and CJK regression test.

Full suite 11708 pass / 0 fail; Python provider suite 146 OK; the schema anti-drift gate re-vendored and green.