v0.13.1

What's Changed

• docs(skill): /ccg — one search section, annotate handoff on empty questions by @tae2089 in #48
• ci: compile and run the postgres-tagged tests by @tae2089 in #62
• test(golden): guard every baseline entry by @tae2089 in #63
• test(golden): make every corpus fail without the structural ranker by @tae2089 in #64
• fix(search): stop paging from hiding a pool that ran out by @tae2089 in #65
• fix(search): index one document per recorded reason by @tae2089 in #66
• fix(search): stop a negative --offset from taking the process down by @tae2089 in #70
• fix(search): give every repository its own file budget in a federated answer by @tae2089 in #71
• fix(search): read a negative offset as zero everywhere, not only at the cut by @tae2089 in #73
• fix(search): cut a query for scoring the way identifiers are indexed by @tae2089 in #74
• feat(search): tell an unannotated repository apart from an unanswerable query by @tae2089 in #76
• test(search): guard that both full-text backends return the best matches first by @tae2089 in #80
• fix(search): order a candidate pool one block at a time so paging holds by @tae2089 in #81
• fix(search): match a Korean question against the reason the index was built from by @tae2089 in #82
• test(db): give every postgres test its own schema so the suite runs in parallel by @tae2089 in #83
• test(search): guard that a suggested skill is one this repository ships by @tae2089 in #84
• fix(mcp): answer a negative offset with a readable sentence on every search surface by @tae2089 in #85
• fix(search): stop one shared character from justifying a candidate by @tae2089 in #86
• docs(search): record what two-rune sub-tokens measured before leaving them alone by @tae2089 in #88
• test(mcp): turn the guides red when they stop matching the MCP contract by @tae2089 in #89
• test(mcp): stop the handler fixture from hiding its own failures by @tae2089 in #90
• refactor(search): give the offset rule its own leaf package by @tae2089 in #91
• test(search): record English-only questions as a decision, not a gap by @tae2089 in #93
• test(search): hold the baseline to the decision recorded in queries.json by @tae2089 in #97
• docs(search): record why the four ranker debts are not yet payable by @tae2089 in #98
• test(search): recapture the golden fixtures from today's source by @tae2089 in #99
• feat(search): index package nodes so a bare package name finds its package by @tae2089 in #101
• fix(search): order tied retrieval rows by identity, not by insertion order by @tae2089 in #107
• test(search): re-judge the query whose judged file was deleted by @tae2089 in #104
• test: judge the package nodes annot and crossref already rank first by @tae2089 in #105
• docs: say what the raw-SQL rule means instead of forbidding what the code does by @tae2089 in #109
• fix(search): break intent ties on identity instead of node id by @tae2089 in #110
• docs(graph): count Identity's fields as five, not four by @tae2089 in #111
• test(search): cover StartLine intent tie breaks by @tae2089 in #114
• test(search): replay raw intent matches in golden eval by @tae2089 in #113
• docs(skill): combine CCG search with source lookup by @tae2089 in #115

Full Changelog: v0.13.0...v0.13.1