[Vzlentin]

optimize ship now improves the whole surface of a run (prompts, profiles, skills, AGENTS.md) from the ship runs that already happened, and keeps an edit only when a blind judge prefers what it produced.

pastRuns("ship")  ->  baseline (item, start, bundle, result)   no rerun unless the head is gone or it did not ship
INV  -> prompts/*.md, agents/*.md edits ┐
DOM  -> skills/*/SKILL.md, AGENTS.md    ├-> Compose -> one overlay + kept|revised|dropped notes -> validate
                                        ┘
run("ship", { from: <baseline run>, overlay })  per item
select(item, diff A, diff B)  blind, stable A/B order, tie = 0
promote when wins/items >= 0.75 and items >= 4  ->  best/, baseline/, overlay.diff, notes.txt
lessons through notes.finish, none when the overlay changes more than one file (confounded)

• run(name, { overlay }): prompts/<n>.md and agents/<n>.md keys replace or remove bundle entries. Every other key goes to the sessions as Spec.overlay (harness 0.6.2), relative to the worktree or absolute, and child runs inherit it. Nothing on disk changes.
• run(name, { from: <runId> }) starts a child with the prompts and profiles of a saved root run, so the overlay is the only difference with the baseline.
• pastRuns(workflow, { runs?, limit? }) is a saved script call over the completed root runs of the store. The agents global gives the profile texts of the run's bundle.
• Validation before any rollout: skill frontmatter, no removed skill that AGENTS.md names, AGENTS.md at most --always-on-max (default the baseline length), and a non-empty overlay. An invalid candidate ships nothing and reports why.
• Breaking: --improve, --trials, prompts/reflect.md and the rubric score are gone. --items <folder> still works, and --runs and --limit (default 8) pick saved runs. TASK.version goes to 18, so runs saved by an earlier version do not resume.