[dbworku]

Both workflows duplicated the plan job, the runner matrix, host info, bench-finish and the summary job. This moves all of it into one place.

• .github/workflows/bench.yml (workflow_call, inputs suite and replicas): the shared pipeline. The plan job validates the suite and replicas and builds the replica list; bench runs the suite once per runner and replica; summary renders the p50 / p90 table. The runner matrix is defined only here.
• .github/actions/typescript-mastra, .github/actions/rust-spotify-player: each suite's phases as a composite action. The commands, pins and phase names are unchanged.
• Entry workflows: each is now just its replicas input and a call to bench.yml.

A single composite action can't hold a job matrix or separate jobs, so the shared part is a reusable workflow. uses: can't take an expression, so bench.yml picks the suite with one if:-gated step per suite.

Docs: the README gains a layout table and steps for adding a runner or a benchmark, and its benchmark table links each suite action. Comments in bin/mark, mise.toml (which still said "Non-Bun") and .github/actionlint.yaml now point at bench.yml.

Failure-path fixes (pre-existing, found in review):

• bench-finish only runs when the harness checkout succeeded, since it is loaded from the harness.
• The summary skips artifacts without host.json instead of crashing the whole table, and counts ok against the requested replicas, so a job that never reported counts as a failure instead of disappearing.
• The summary no longer runs when plan rejects the inputs.

actionlint, zizmor and shellcheck are clean, and the summary was tested locally against fixtures for these failure paths. Not yet run in Actions.