fix(knowledge): index only searchable model

[ohdearquant]

AI-assisted contribution: Codex prepared this change and PR description.

Summary

• make knowledge.index embed and persist atoms with only the configured default model, matching the model selected by every current knowledge search, suggest, compose, and section-retrieval path
• stop building, invoking, and writing unread secondary knowledge embeddings while retaining multi-model behavior for substrates that actually fan out retrieval
• await the default-model indexing batch in the dispatch task so its work remains represented by the request's usage accounting
• add a two-engine regression proving the default provider is called, the secondary provider is not called, primary vectors are written, and no secondary vector table is initialized
• reconcile operations guidance and ADRs around the current default-model contract and the symmetry required before future knowledge multi-model fanout

Closes #1513.

Test plan

• cargo test --manifest-path crates/Cargo.toml --workspace
• cargo check --manifest-path crates/Cargo.toml --workspace
• cargo clippy --manifest-path crates/Cargo.toml --workspace --all-targets -- -D warnings
• cargo fmt --manifest-path crates/Cargo.toml --all -- --check
• RUSTDOCFLAGS=-Dwarnings cargo doc --manifest-path crates/Cargo.toml --workspace --no-deps

All five gates passed against independently reviewed SHA 74669ec50052573bbd40defa610f7644b5278869 on unchanged main base c32ea3541063ca5626058dd52e9d5f419f7f2181 before this draft was opened.

ADR

• amend ADR-031 to require readable-model symmetry before a writer fans out across embedding models
• align ADR-035, ADR-047, ADR-051, and ADR-103 with the knowledge pack's current default-model indexing and retrieval contract

AI-assisted contribution checklist

• Every claim in this PR description matches the actual diff
• Any agent-authored comment / PR body starts with an attribution line
• The regression uses fake providers and performs no model download
• Full workspace behavior, lint, formatting, and documentation gates passed

Out of scope

• adding multi-model knowledge retrieval
• migrating section embeddings to model-keyed storage (Store knowledge-section vectors per embedding model #1511)
• changing multi-model indexing or retrieval behavior in memory, communication, or generic KG paths