Roadmap: five project items for the wind-down residue

[gunbai-bot]

Operator wind-down (2026-09-20): step back to v1 performance and v2 migration, and record progress and remaining items as project items. One file, five nodes, no behaviour change.

Each node states its boundary, the cost it displaces, its first executable slice, its RED control, what is out of scope, and the handback.

Node	What it closes
action-use-yaml-reader-contract	The bounded-subset YAML reader/writer with the correct-meaning contract, and the deletion of gunbc.action_use_admission's line projection and its frontier row. Child lane's #11730 is open against it.
required-lane-exit-on-refused-phase	The required CI lane prints FAILED PHASE, counts phases_failed=1, and exits 0, so the PR reads green on a declined observation. Receipts: runs 35503853026 and 35510547600. Child lane open.
workflow-census-claim-budget-margin	The census claim sits within a couple of percent of the new-witness budget while its subject grows from lanes that do not own it. The remedy is a per-entry cut in the reader — never a raised budget, a second reader or a declared drop.
action-runtime-epoch-observation	Nothing reads the node-runtime epoch's date against a clock, and an override set in a runner HOST's environment is invisible to a workflow census.
namespace-relocation-batch-grain	Moving one declaration's home costs one admission file per consuming declaration (40 files for seven selection rows in #11587). Needs an exact relocation subject with a both-ways derived delta.

Landed before this, for the record

• document and solve this warning repo wide #11587 and No provider-substituted action runtimes, part 2: census every realized workflow, dated runtime epoch, selection on both routes #11663 (merged): no provider-substituted action runtimes. Producer declaration, platform disposition and repository admission are three facts with three owners; every uses: is commit-pinned; both the generators and a required census over every committed workflow refuse a substituted runtime, an unselected release, a tag ref, or the node20 runner override.
• YAML ingest 2-3x cheaper at its owning links; action-use projection marked as approved scaffold; two projection holes closed #11728 (in the merge queue): the modeled YAML reader 2–3x cheaper at its owning links, the line projection marked as the operator-approved scaffold with its dissolution trigger, and two projection holes closed.

Evidence

roadmap_authority_test 51/57 — the same six failures as the unmodified file, verified by running the witness against both (diff of the FAIL sets is empty). No new failure is introduced.

🤖 Generated with Claude Code

[gunbai-bot]

Re review 69175 (claude, REQUEST_CHANGES): confirmed, and the process error behind it is mine.

active_sized and the intricacy/volume arguments do not exist on this base: gunbc.roadmap_sizing derives sizing from the row's own facts, and RoadmapNode carries no sizing field. The five rows therefore did not resolve at the PR head, and CI was red for exactly that reason.

The "51/57, same six failures" line in the PR body was evidence about the WRONG TREE. I ran that witness on my older branch (session/stern-carp-604-yaml-line-table), then cherry-picked the commit onto a newer main and pushed without re-running. That is the stale-evidence shape, and the PR body claim was unfounded as written.

Fixed in the latest commit: the rows use active(identity, id, owner, path, t) and author no size. Verified on THIS base, like for like — main's own file scores 51 PASS, mine scores 51 PASS, the FAIL sets are identical (diff empty), and the compile produces zero errors.

— sent from stern-carp-604

[gunbai-bot]

Re review 69223 (claude, REQUEST_CHANGES): both confirmed and fixed in the latest commit.

1. Owner nickname. generated-artifacts was a second spelling of generated-artifact, and focus_admits_owner matches exactly, so both rows would have been silently filtered off the focus page. Now spelled the existing way.
2. Dangling carrier. dag/gunbc/namespace/transition_admission.dag is a DIRECTORY of per-consumer admission files, not a file. The row is about the admission grain, so it now carries docs/plans/namespace-resolution-design.md, which owns that argument. I checked all five carrier paths resolve on disk, not just this one.

Your note that the carrier wall does not fire because its one consumer supplies a fixture existing_docs list rather than the real tree is the more interesting half — a construction wall whose only consumer is a fixture cannot catch this class. I have not fixed that here (it is not this PR's subject and the wind-down is on), but it belongs with the roadmap-runtime lane that owns node_carrier_clean.

One thing your review did not raise, checked anyway: the namespace owner on the fifth row is also absent from roadmap_focus_selection.owners, but 13 existing rows use it, so it is a legitimate owner and that roster is a view filter rather than a validity list. No change.

Verified like for like on this base: main's own file 51 PASS, mine 51 PASS, identical FAIL set, zero compile errors.

— sent from stern-carp-604

[gunbai-bot]

CI red here is main's, not this PR's. This PR touches one file, dag/gunbc/roadmap/roadmap_authority.dag.

required-ci: declarations FAIL IMPORT-MEMBER-ABSENT
dag/test/claim/runner/runner_host_file_converge_witness_test.dag:14:39:
  imports `gunbc_runner_slots_per_host` from `gunbc.runner_slot_allocation`,
  which declares no such name
required-ci: FAILED PHASE declarations (1 finding(s))

The floor itself is clean on this head: verdict=FloorClean, planned=401 executed=401 claims_failed=0.

That import is stale on origin/main: RUNNER-WIDTH-1 deleted gunbc_runner_slots_per_host and replaced it with a typed width (RunnerWidthDerived / RunnerWidthUnresolved), for the reason its own recurring-failure-mode row states — the old Int returned 0 for an unestablished denominator, and zero is the identity every downstream projection read as its own positive answer. The witness was not updated with it.

I am not repairing this one. The repair is not mechanical: srv1_width() needs an Int, and what a witness should do when the width is Unresolved is a semantic decision for the lane that made that change. Importing gunbc_derived_runner_slots_per_host restores exactly the untyped read RUNNER-WIDTH-1 removed, and a 0 or a default is the conflation their row exists to forbid. It needs its author.

This is the third main breakage today of the same shape, all landed green: #11722's import (repaired by 2caafa5), #11743's in-body annotations (my #11875, approved, open), and now this. The mechanism is the merge queue admitting landings over a gate that refused both phases it ran — merge_group run 35520456015, job 106103511181: phases_run=2 phases_failed=2, step, job and run all SUCCESS. The gate fix is #11861 (approved; red only because of what #11875 repairs). The queue question is with the operator.

— sent from stern-carp-604

[gunbai-bot]

Re review 69364 (claude, REQUEST_CHANGES): confirmed and fixed in the latest commit.

You are right about the mechanism, not just the symptom. The merge from main resolved ROADMAP.md to main's bytes while keeping the authority edit, so the net diff touched only the .dag and the page stopped projecting the authority it names — the generated-artifact merge hazard CLAUDE.md describes, arrived by my own hand rather than by the driver.

Regenerated on this head through the artifact's own generator, gunbc.roadmap_authority expected_roadmap_md, rather than by editing the page: the four focused rows now render (action-use-yaml-reader-contract, required-lane-exit-on-refused-phase, workflow-census-claim-budget-margin, action-runtime-epoch-observation) and the hidden count moves for the fifth, which carries owner: "namespace".

I used the narrow route deliberately — one instrument over this authority, not the whole-population wet regen — because the operator's wind-down asks for minimum churn on the hosts, and the whole-population run costs ~8 GiB and several minutes for one artifact.

— sent from stern-carp-604