Sync Boatstack from Intelligence Flow Labs @ 60f76062f721

[operator-stack-publisher]

What this sync releases

Corrections for a published-but-open slice route to that slice in place, not the active one

When a multi-slice delivery publishes an early slice and the build pointer advances, that slice is not finished: its pull request is still open and may need a CI fix or a review follow-up before it merges. Boatstack's actuators already knew this — resolveAddressableSlice treats the addressable set as {the active slice} ∪ {published slices whose PR is not terminal}, so record-delivery-gate --slice and publish-pr can re-gate and update a published-open slice in place. But the advisor, router, and safety layers were left deciding purely on the whole-delivery pointer (ActiveIndex < len(Slices)): whenever any later slice was still unbuilt they treated the correction as belonging to the active slice. A fix pushed on the published slice's own branch was mis-routed — recovery-status named the active slice and prescribed repair_active, record-change charged the active slice's repair budget and reset it to BUILD, and the publication-bypass finding reported the fix branch as relation=unrelated against the active slice. The published slice's open PR was left with no clean forward move, and the active slice's build loop was corrupted by a correction that was never about it. This is the slice-granularity instance of advisor–actuator addressability drift: one domain concept — "which slice may still be corrected" — with two implementations, only one of which had been migrated to per-slice addressability.

The advisor layer now consumes the same addressability the actuators do. A single branch-keyed resolver, resolveAddressableSliceByBranch, returns the addressable slice (active or published-open with a non-terminal PR) that a correction branch owns, and the router, recorder, and safety finding all route through it. A correction reported from a published-open earlier slice's branch now resolves to that slice: recovery-status reports PUBLISHED_OPEN and a repair_published_slice next-operation that directs the operator to re-gate and publish-pr --action update the slice in place; record-change (now with an optional --slice, falling back to the correction branch) binds the observation to that slice and drives an in-place re-gate that preserves its open-PR identity, without touching the active slice's status, repair budget, or mode; and the publication-bypass finding names the published slice and reads the branch as current_branch. The active-slice path is unchanged for the ordinary case, and a fully-published (terminal-PR) delivery keeps its existing draft_corrective_child behavior.

The regression suite pins the class shut: from a published-open slice's branch it asserts the router routes in place, record-change binds the published slice while leaving the active slice untouched (both via --slice and via branch resolution), the safety finding names the published slice, and — the anti-drift invariant — the advisor's by-branch resolver and the actuator's by-id resolver select the same slice, so the two layers cannot silently diverge again.

Projection provenance

Generated from operatorstack/intelligence-flow@60f76062f72185efddad1c22f6c1fc088aff0005.
Review provenance, tests, and examples before merging.