Card Headline Truth

This sprint was the loader slice. It had to extend the pattern format and reference dataset without changing live card rendering yet, which made the boundary narrow enough for a fresh worker. Its proof was structural: accepted step kinds preserve a non-empty literal title, and invalid title data fails before execution.

This sprint connected the new pattern field to card creation. The scope was deliberately runtime-local: choose among declared title, task title, declared prompt or command, generated headline input, and the stage fallback. That kept it independent from the broader pattern-file sweep that followed.

This sprint was the data propagation pass for the title contract. It did not need to change engine behavior, because the previous sprint had already made titles effective. Its risk was omission, so the useful boundary was a check that fails when a bundled executable step leaves the title empty.

This was the publication closeout for the first milestone. The work was thin by design: carry the title-key contract into the reference-card documents after implementation and pattern data had passed. Its boundary was docs synchronization, so no runtime stress belongs to this sprint.

This sprint isolated the qualifier problem from the rest of the card text. The worker only had to make gate cards use composed task paths and remove stale hardcoded prefixes, while preserving explicit pre-merge and post-merge facts from their callers. It passed at the sprint level, but the milestone later needed a retry, so this slice fed a harder shared-runtime integration.

This sprint corrected a false stage label without deleting the historical stage value. The worker had to separate bundled plan-change review steps from real replan work, then keep role, harness, and effort selection aligned with review. That made the slice semantic rather than mechanical, but narrow enough to prove through pattern execution paths.

This sprint closed the milestone by making the policy vocabulary visible in the generated pattern material. The implementation work it depended on had already landed, so the boundary was synchronization across reference snapshots. It remained green because it did not reopen runtime behavior.

This sprint was the format admission step for card summaries. It did not need to run summarizers; it only had to make `summarize` legal where the later runtime could consume it. The isolated parser and test boundary let the worker prove the key without taking on finish-card behavior.

This sprint attached the accepted key to the existing summarizer path. The hard requirement was failure tolerance: a missing, failed, or empty generated summary must leave the ordinary finish summary in place. That gave the worker a precise runtime contract and avoided a new history schema.

This was the documentation and publication pass for the summary key. The runtime contract already existed, so the worker only needed to make the repository and website documents agree. The run facts show no local stress for this docs-sized slice.

This sprint turned the plan's verification requirement into a fresh run audit. Unlike the earlier sprints, it had to inspect the rendered record rather than add one isolated key. The audit found card-truth problems in engine-emitted cards, and the sprint closed only after fixes and re-audits showed the in-scope findings gone.