Release v7.103.1

Changed

• The Screenplay skills now warn about the traps found in SoloCRM run 2: generated identifiers for new entities and the for pin in specifications, a STAGE-ESM-017 checklist with bisecting through render --check, where hand-written queries go, managed-artifact refusals, PLAY0269 not being evidence of rendering, application titles, footers and action labels, a run-and-verify recipe, and running the specs before calling a model done (#854)
• The Stage admission reference now says that a projection may map onto the identifier when the mapping restates the key: the key's own event property, $eventSourceId for an event-source key, or the same literal.
• The ai-verification test suite no longer runs the full corpus verification eight times. Frontmatter variants are checked directly against the real skill, and three end-to-end runs (accepted corpus, a rejected trigger, CI without a base) run side by side. (#875)
• The agent-harness entrypoint fixtures run through one runner, Source/AgentHarnesses/run-entrypoint-fixtures.sh, that runs independent fixtures concurrently and keeps every exit code. On a non-Linux developer machine the four Linux-only fixtures report SKIPPED with the reason; on Linux and in CI they always run. (#876)