feat(search): add bounded archive discovery with eligible replies

[leemour]

Natural questions require words that the existing strict matcher cannot find together, and short replies often lack the query's words entirely. Add explicit archive discovery through CLI search messages --discover, MCP discover=true and SDK SearchQuery.discover. Partial lexical retrieval uses existing SQLite FTS/BM25, ranks bounded evidence with stemmed term coverage and reciprocal retrieval rank, and recovers eligible direct replies. A reply inherits its parent's lexical rank and comes first on an equal score. No neural assets, full-archive snapshot or automatic conversation rebuild is used.

Hard account/chat/sender/date/only constraints apply to both parent and reply. Explicit Boolean syntax, phrases, wildcards, AST, exact and newest requests retain strict behavior. Matched/missing terms, reply-parent provenance, candidate depth and truncation make partial evidence visible. Discovery searches the archive, refuses legacy/regex/server-only modes, and preserves requested surrounding/thread context. Saved runs keep the additive option. Plain keyword discovery also preserves used prefix/correction provenance.

Migration 29 adds a derived live-reply index, reserved by merged PR #817; no message data changes. Rewrite only selection into indexed message-key lookups, avoiding a correlated JSON test over every archive row. Retrieve at most 300 lexical candidates and 100 proposed replies, then keep at most 300 scoring candidates under existing body/time budgets. A full-shortlist regression test ensures replies can enter the final pool.

The original one-default proposal remains unqualified by its approved quality gates. This is an additive opt-in; strict CLI/MCP defaults and no-language SDK compatibility remain unchanged. The source handoff and query guide describe this boundary. Synthetic research, original artifacts, the ranking matrix and fresh/scaling reports live in WireCatLabs/cli-testing PR #5. The freshly authored fixture is not an independent human/real-world evaluation; rank one and paraphrases remain limitations. Scores are not answer confidence.

Validation: lint, typecheck, 2790 passing tests (2 existing skips), coverage floors, documentation checks and Bun smoke pass. Isolated SDK tests cover scope/account/deletion, full candidate pools, syntax/AST/flags, contexts, typo provenance, cancellation and bounds. Telegram and MAX offline adoption tests are being verified against a packed build before the shared release and exact consumer pin bumps. No real account/store/messenger operation was performed.