What it means
Nobody keeps a list of what to benchmark. Each executable capability with a required policy is a target directly and on every transport its exposure declares, with the cases its input type provides; the transports' own operations (a cold majordomus mcp process, initialize, ping, tools/list, resources/list, resources/read, GET /, /openapi.json, /docs) are targets declared once. majordomus bench coverage reports covered over required, where required is computed from the registry; a capability whose input type produces no case for the repository at hand is missing, and capabilities validate, bench coverage --check and CI fail on it. A waiver is a typed reason on the descriptor, reported and never counted as covered.
How it works
apps/majordomus-cli/src/bench/projection.rs derives the targets from the registry and the case providers; bench/coverage.rs computes the requirements and the tallies; bench/runner.rs times a target through the executor, over a real loopback socket with the input bound as the route binds it, and through a real majordomus mcp --standalone child; bench/results.rs and bench/baseline.rs hold the versioned result document, the per-platform baseline and the regression policy read from .ai/repo/benchmarks/rust/policy.yaml; commands/bench.rs is the command line.
How to see it
apps/majordomus-cli/target/debug/majordomus bench coverage # per transport and in total: covered / required, missing 0, waived 0
apps/majordomus-cli/target/debug/majordomus bench --profile quick # every target, slowest p95 first; written under .ai/local/benchmarks/
apps/majordomus-cli/target/debug/majordomus bench objects.search --transport mcp --profile full --format json
apps/majordomus-cli/target/debug/majordomus bench --profile ci --check # against .ai/repo/benchmarks/rust/baseline.<platform>.json, under the policyWhat it does not cover
Wall-clock numbers depend on the machine: a baseline is compared only on its own platform, and CI runners without a committed baseline compare nothing. The process-cold target is noisy by nature; the policy's absolute floor and per-metric thresholds are data a person tunes with evidence. The MCP transport measures a warm process for a cached capability (cold is not observable from outside); the direct transport reports cold and warm.
Why it exists
"Representative endpoints only" is how the slow one escapes; the operator asked for every operation, with the denominator generated. project.rust-benchmark-coverage is the rule, ADR 0004 the decision. apps/majordomus-cli/tests/bench.rs proves the derivation, that an exposure added or removed adds or removes the requirement and the target with no other edit, that the three runners time real transports, and that a check reports every line and never attaches a stale baseline to a renamed target; test/cases/91_canonical_architecture.sh recomputes the denominator from the live registry and compares it with the coverage document.