v1.108.145 — coverage contract on absence claims + scorer-pinned calibration

An absent verdict backed only by scan counts can lie by omission. Every full discovery walk now persists a coverage block into the index metadata (files indexed, skip counts by reason — unsupported extension, oversize, binary, secret, cap-dropped — plus zero-symbol file count), and absent/degraded verdicts disclose it alongside index generation and included scopes, across search_symbols, search_text, get_ranked_context, and the file/symbol read tools. An index that predates the contract omits the block: coverage-unknown is never presented as "nothing was excluded".

Verdicts also carry a scorer version pin, and benchmarks/calibration/planted_queries.json records planted positive/negative query rates re-measured live in CI — a scorer change without a re-measured artifact fails the build. Full reliability curves (ECE/Brier per query class) remain the stated end state.

The published retrieval-verdict JSON Schema documents both additions. No INDEX_VERSION bump; no tool-count or input-schema change.

Origin: community feedback on the retrieval-verdict article. Suite parity: jdatamunch-mcp v1.20.0 (shipped), jdocmunch-mcp in flight.