diff --git a/devlog/_plan/260914_round23_delivery_roadmap/000_roadmap.md b/devlog/_plan/260914_round23_delivery_roadmap/000_roadmap.md
new file mode 100644
index 0000000000..5fb5389c58
--- /dev/null
+++ b/devlog/_plan/260914_round23_delivery_roadmap/000_roadmap.md
@@ -0,0 +1,64 @@
+# 260914 — Round 2/3 delivery roadmap
+
+## Where this starts
+
+The 2026-09-14 triage delivery loop closed with #4519, #4534, #4535, #4536,
+#4539, #4555, #4562 and #4563 on dev. What it did not do is reduce the open
+surface: 63 pull requests and 63 issues are still open, and a large share of the
+pull requests are contributor work that is already green and simply never got
+landed. This unit plans two more delivery rounds against that backlog.
+
+## Objective
+
+Rebuild the lane roadmap, run two merge rounds, and finish with a closure sweep
+that closes at least 10 issues and at least 10 pull requests with written
+reasons. Every close names either a merge commit or the change that supersedes
+it.
+
+## Delivery constraints carried from the session
+
+- Lanes are **worktree Codex threads**, one per branch. Subagents inside a lane
+  are that lane's workers; they never own a branch of their own.
+- Lane orchestrators split between `anthropic/claude-opus-5` and `kimi/k3[1m]`.
+- Inside a lane, implementation is delegated to `devin/swe-2` and
+  `xai/grok-4.6` subagents at roughly a 2:3 ratio.
+- Each lane runs `cxc-loop` itself: its own goal, its own PABCD cycle.
+- **No local suite.** `bun run test`, `bun run typecheck` and `bun install` are
+  not run locally. Proof is hosted CI at the exact final head, nothing else.
+- Pushes use `--no-verify`.
+- Merges into dev are admin squash merges, taken on maintainer judgment under
+  the MAINTAINERS.md single-maintainer dev integration policy, with exact-head
+  CI recorded.
+- Landing another author's work carries a `Co-authored-by` trailer in a branch
+  commit so it survives the squash.
+
+## Round shape
+
+Two merge rounds, then one closure sweep.
+
+| Round | Work | Merge target |
+|---|---|---|
+| 1 (wp2) | 5 lane threads on issues that have no pull request, plus review-and-land of the already-green contributor queue | dev |
+| 2 (wp3) | 4 lane threads on the second issue cluster, plus the second contributor batch | dev |
+| Sweep (wp4) | Close resolved issues with merge references, close superseded and abandoned pull requests with reasons | — |
+
+## Acceptance criteria
+
+1. This roadmap unit exists with per-lane issue assignment, dispatch models and
+   acceptance criteria.
+2. Round 1 lanes were dispatched as separate worktree threads and their pull
+   requests reached green hosted CI at their exact final head.
+3. Round 1 pull requests were admin squash-merged into dev with recorded merge
+   commits.
+4. Round 2 lanes were dispatched and squash-merged into dev the same way.
+5. At least 10 issues are closed with a merge-commit reference or a written
+   supersession reason.
+6. At least 10 pull requests are closed — merged, or closed with a written reason
+   naming the superseding change.
+
+## What would make this fail
+
+The backlog is contributor work, so the failure mode is not "no code lands", it
+is "code lands that nobody reviewed". A green check on a three-day-old head is
+not evidence about the head being merged. Every merge in this unit re-reads the
+diff at the head it is about to squash, and re-reads CI at that same SHA.
diff --git a/devlog/_plan/260914_round23_delivery_roadmap/010_lane_assignments.md b/devlog/_plan/260914_round23_delivery_roadmap/010_lane_assignments.md
new file mode 100644
index 0000000000..82a69602b8
--- /dev/null
+++ b/devlog/_plan/260914_round23_delivery_roadmap/010_lane_assignments.md
@@ -0,0 +1,141 @@
+# 010 — Lane assignments
+
+Each lane is one worktree Codex thread on its own branch. The orchestrator model
+is named per lane; inside the lane the implementation work goes to devin/swe-2
+and xai/grok-4.6 subagents at roughly a 2:3 ratio. Every lane runs cxc-loop,
+pushes with --no-verify, opens one pull request against dev, and runs no local
+suite.
+
+This split is the second revision. The first one was audited and failed: two
+round-1 lanes both owned src/server/responses/core.ts, two round-2 lanes both
+owned the Models dashboard, and src/codex/catalog/effort.ts had two claimants
+across rounds. Lanes in the same round are now disjoint by file, and the only
+remaining overlaps are sequential, where the later round branches from dev after
+the earlier one has landed. The write-scope exclusions below are the mechanism;
+they are part of each lane's dispatch, not advice.
+
+## Round 1 lanes — branch from current dev
+
+| Lane | Orchestrator | Issues | Carried | Write scope |
+|---|---|---|---|---|
+| R1-L1 catalog normalization | claude-opus-5 | #4570, #4505 | — | src/codex/catalog/parsing.ts and catalog normalization; **not** catalog/effort.ts |
+| R1-L2 pool routing and prompt cache | kimi/k3 | #4546, #4550 | — | src/codex/routing.ts, pool and session affinity; **not** src/providers/devin*, src/providers/antigravity*, src/server/responses/* |
+| R1-L3 cursor policy and errors | claude-opus-5 | #4508, #4542 | #4509 (HeiTuz), #4544 (001005HS) | src/adapters/cursor/* |
+| R1-L4 responses terminal, reasoning payload, media | claude-opus-5 | #4469, #4311, #4312, #4532 | #4549 (jiaoyun286) | src/server/responses/*, src/responses/*, src/adapters/anthropic* |
+| R1-L5 provider account lifecycle edges | kimi/k3 | #4503, #3781 | — | src/providers/devin*, src/providers/antigravity*; **not** the pool kernel |
+
+### What each lane owns
+
+**R1-L1** — #4570 is a live-catalog regression where
+zhipu-bigmodel-coding/glm-5.3-flash loses its contextWindow on 2.54.0, so the
+model arrives with no context budget. #4505 is the neighbouring modality gap:
+DeepSeek V4.1 Flash and GLM-5.3 Flash declare no image input, which disables
+combo image routing on OpenCode Go and CommandCode. Both are parsing and
+normalization. The reasoning ladder is explicitly not this lane's, because
+R2-L6 owns effort.ts.
+
+**R1-L2** — #4546 is the expensive one: pool routing rotates accounts mid-thread,
+which destroys the prompt-cache prefix and produces a 10x-50x token burn above
+the 80% usage threshold. #4550 is the same subsystem from the other side, an
+existing Codex CLI thread that bypasses the pool over a direct WebSocket while
+status still claims opencodex-local.
+
+**R1-L3** — #4508 loses the Cursor data-policy "action required" signal into a
+bare failed_precondition, and #4542 makes the native-exec refusal name
+shell_command/exec_command even when the catalog has no shell tool, which makes
+kimi-k3 abandon the turn. #4509 and #4544 are contributor attempts at exactly
+this surface; both are carried with attribution rather than reimplemented blind.
+
+**R1-L4** — this is the merged Responses lane. #4469 is reasoning
+encrypted_content never reaching the caller that asked for it; #4311 is paginated
+Codex history silently stopping projection after provider metadata relabeling;
+#4312 reports an Anthropic content_filter terminal as a 502 so Codex retries
+something that can never succeed; #4532 is dynamic image downscaling on append
+busting the prompt prefix cache, which #4549 fixes. They were two lanes until the
+audit showed #4549 already diffs core.ts, encrypted-payload.ts, compact.ts and
+collaboration.ts — the same files #4469 needs. One lane, one merge.
+
+**R1-L5** — #4503 is the Devin provider's detached credential rekey leaving a
+host-selection window, and #3781 is the Antigravity quota-refresh failure with
+missing canonical Fake-IP handling. Both are vendor-local and stay out of the
+pool kernel that R1-L2 owns.
+
+## Round 2 lanes — branch from dev after round 1 lands
+
+Cutting these after round 1 is what makes the remaining overlaps safe. R2-L6
+needs src/codex/catalog/parsing.ts, which R1-L1 also changes, and #4461 edits it
+too.
+
+Round 2 runs in two waves, and inside wave A exactly one lane owns the config
+schema. That rule is the thing that took three audit passes to get right: any
+lane that adds a user-facing setting has to edit src/config.ts and
+src/types/config.ts, so "disjoint" cannot be expressed in feature terms alone.
+L8 is the schema owner for wave A; L6 and L7 land behavior only, and if either
+one genuinely needs a new setting it files a follow-up instead of editing the
+shared schema.
+
+The waves exist for the same reason. Wave A is L6, L7 and L8, which
+are disjoint from each other. Wave B is L9 alone, branched from dev after L8
+lands. The second audit found why: #3630 and #3377 are config-schema changes in
+src/config.ts and src/types/config.ts, and L9's carried #4042 already edits both
+of those files plus src/server/background-lifecycle.ts, which is exactly where a
+refresh timer would register. Naming L8's scope as catalog-refresh-status.ts and
+convergence.ts did not remove that collision, it only hid it. Sequencing does
+remove it.
+
+| Lane | Orchestrator | Issues | Carried | Write scope |
+|---|---|---|---|---|
+| R2-L6 codex runtime and Windows probe (wave A) | claude-opus-5 | #4204, #4458 | #4461 (S0RYUASUKA) | src/codex/runtime.ts, catalog/effort.ts, catalog/bundled.ts, catalog/parsing.ts; **no new config-schema fields** |
+| R2-L7 web-search bridge (wave A) | kimi/k3 | #4429 residual, #2730 | — | src/web-search/*, src/server/search.ts; **no new config-schema fields** |
+| R2-L8 catalog auto-refresh backend (wave A) | claude-opus-5 | #3630, #3377 | — | src/codex/catalog-refresh-status.ts, convergence.ts, src/config.ts, src/types/config.ts, src/server/background-lifecycle.ts; **not** gui/, **not** management/config-routes.ts |
+| R2-L9 dashboard model and usage surface (wave B, branches after L8 lands) | kimi/k3 | #4175, #4209 | #4193 (chilung-cgu), #4042 (Vocllum) | gui/src/pages/Models.tsx, gui/src/pages/Usage.tsx, src/server/management/config-routes.ts, and the config-schema files only as rebased on top of L8 |
+
+**R2-L6** — #4204 has a stale persisted CLI 0.135.0 stripping max and ultra while
+Codex Desktop runs 0.153.4, and #4458 is the Windows prompt probe missing the
+Codex App runtime and its base prompt source. Both are runtime resolution, and
+effort.ts has exactly one owner in the whole unit.
+
+#2279 was in this lane and is deferred out of the unit. It asks for a per-model
+setting that suppresses synthetic max while retaining ultra, which is a
+config-schema field, and the third audit pass showed that putting a schema field
+in L6 collides with L8 no matter how the rest of the scope is drawn. It is a real
+ask and it stays open; it just does not fit a lane that has to run beside a
+schema owner.
+
+**R2-L7** — #4429 is only partly landed. #4515 (cb2e15ba6f) shipped the
+passthrough backends and said so, but the residual is mixed-tool continuation:
+a key-auth Responses gateway (Kimi K3) still echoes hosted web_search back as a
+client function_call, and webSearchBridge stays Ollama-only. The lane closes the
+residual, not the whole issue as if nothing had landed. #2730 is the related ask
+to let /v1/alpha/search use a configured backend without ChatGPT forward auth.
+
+**R2-L8** — #3630 wants periodic catalog auto-refresh so newly released models
+appear without a manual ocx sync. #3377 wants per-model capability declarations
+for text-only, context tier, and video processing mode. Backend only; the
+dashboard side of the same story belongs to R2-L9.
+
+**R2-L9** — #4175 is the Dashboard toggle for Fast selector rows, which #4193
+already implements. #4209 wants the dashboard to distinguish a hub model change
+that is saved from one that is synced from one that is actually active on a given
+client. #4042 adds a configurable usage history size limit and has gone stale.
+
+## Contributor merge queue
+
+These need review plus an admin squash, not new implementation. The reviewers
+found that every one of them was sitting with Cross-platform CI and React Doctor
+in action_required, which is the fork-workflow approval gate — so none of them
+had ever produced exact-head suite proof. Approving those runs at the current
+head is the first step for each, and the merge waits on the result.
+
+- Batch 1 (round 1): #4565, #4452, #4451, #4383, #4139, #4517, #4298, #4071,
+  #4033
+- Batch 2 (round 2): #3952, #3748, #3742, #4224, #4265, #4199, #4177, #3833,
+  #4178, #4566, #4564, #4568, #4569
+- Held: #4309 is 288 commits behind dev and its provider-count assertions cannot
+  be trusted until it is rebased.
+
+Draft state is a gate artifact, not an author objection: the contributor gate
+opens fork pull requests as drafts and holds them until a four-box checklist is
+ticked, and the local-CI box is an attestation a fork author cannot satisfy
+because fork contributors cannot start repository CI. Marking such a pull request
+ready before an admin squash is a maintainer action, and it is recorded as one.
diff --git a/devlog/_plan/260914_round23_delivery_roadmap/020_merge_and_closure_policy.md b/devlog/_plan/260914_round23_delivery_roadmap/020_merge_and_closure_policy.md
new file mode 100644
index 0000000000..04c569e647
--- /dev/null
+++ b/devlog/_plan/260914_round23_delivery_roadmap/020_merge_and_closure_policy.md
@@ -0,0 +1,52 @@
+# 020 — Merge and closure policy for this unit
+
+## Merge policy
+
+Every merge in this unit is an admin squash into dev. The authority is the
+single-maintainer dev integration clause in MAINTAINERS.md: a maintainer with
+GitHub maintain or admin access may integrate through a pull request without a
+second approval, including their own, provided the decision and the exact-head CI
+evidence are recorded. That clause covers dev only. main and preview are
+untouched here, and no direct push to a protected branch happens at any point —
+the bypass is pull-request-only, which --no-verify does not change.
+
+Squash is the merge method for the whole unit. What gets verified is the round's
+final tip after the squash, not each intermediate branch state, because a squash
+collapses the branch into one commit on dev and the only thing dev ever sees is
+that commit.
+
+Before any squash the lane's diff is re-read at the head about to be merged, and
+the check rollup is read at that same SHA. A green run on an older head is
+evidence about the older head.
+
+## What a lane must deliver
+
+1. One branch, one pull request against dev, filled to the repository pull
+   request template (Summary, Verification, Checklist).
+2. Closes #N lines for every issue the lane resolves. GitHub will not auto-close
+   them because these target dev rather than main, so the sweep closes them by
+   hand with the merge reference.
+3. A Co-authored-by trailer in a branch commit for every carried contributor pull
+   request, so it survives the squash and reaches the contributor graph. Prose
+   naming the author is not equivalent; CREDITS.md exists because that mistake
+   was made 27 times.
+4. Verification stated honestly: hosted CI at the exact head. Local suite runs
+   are labelled NOT RUN because they were not run.
+
+## Closure policy
+
+An issue closes when a merged commit on dev resolves it, and the closing comment
+names that commit. An issue also closes when it is superseded, and then the
+comment names what superseded it and why the original ask is satisfied.
+
+A pull request closes one of three ways in this unit:
+
+- **Merged.** The ordinary case for the contributor merge queue.
+- **Superseded.** A lane landed the same behavior; the closing comment names the
+  merge commit and the carried attribution, so the author can see where their
+  work went.
+- **Abandoned against current dev.** The branch conflicts, the approach no longer
+  matches the subsystem it targets, and reviving it would cost more than
+  reimplementing. The comment says that plainly and leaves the door open.
+
+No pull request is closed merely for being old, and none is closed silently.
diff --git a/devlog/_plan/260914_round23_delivery_roadmap/030_round1_dispatch.md b/devlog/_plan/260914_round23_delivery_roadmap/030_round1_dispatch.md
new file mode 100644
index 0000000000..be9e277888
--- /dev/null
+++ b/devlog/_plan/260914_round23_delivery_roadmap/030_round1_dispatch.md
@@ -0,0 +1,44 @@
+# 030 — Round 1 dispatch record
+
+Five worktree Codex threads were created for round 1, one per lane, each with its
+own checkout and branch. Each was told to run cxc-loop itself, to delegate
+implementation to devin/swe-2 and xai/grok-4.6 subagents at roughly a 2:3 ratio,
+to run no local suite, to push with --no-verify, and to open one pull request
+against dev without merging it.
+
+| Lane | Orchestrator | Branch | Issues |
+|---|---|---|---|
+| R1-L1 catalog normalization | anthropic/claude-opus-5 | codex/260914-l1-catalog-normalization | #4570, #4505 |
+| R1-L2 pool routing and prompt cache | kimi/k3[1m] | codex/260914-l2-pool-routing-cache | #4546, #4550 |
+| R1-L3 cursor policy and errors | anthropic/claude-opus-5 | codex/260914-l3-cursor-policy-errors | #4508, #4542 |
+| R1-L4 responses terminal and media | anthropic/claude-opus-5 | codex/260914-l4-responses-media | #4469, #4311, #4312, #4532 |
+| R1-L5 provider account edges | kimi/k3[1m] | codex/260914-l5-provider-account-edges | #4503, #3781 |
+
+## Contributor merge queue, batch 1
+
+The review pass found the thing that had actually been blocking this queue.
+Every fork pull request in it was sitting with Cross-platform CI and React Doctor
+in `action_required` — the fork-workflow approval gate — so the suite had never
+run at those heads. The green checks visible on each pull request were the
+hygiene, labeler and target gates only. Nine runs were approved at the exact
+current head:
+
+| PR | Head | Approved run |
+|---|---|---|
+| #4565 | 449a692f4 | 34794874697 |
+| #4452 | 4b213ef2b | 34796097782 |
+| #4451 | daf735d67 | 34794900986 |
+| #4383 | 713fbe66e | 34794898049 |
+| #4139 | 2a8ca25cb | 34794890662 |
+| #4517 | 3a45afe6a | 34795340237 |
+| #4298 | cdbac2727 | 34778679348 |
+| #4071 | 78a9e097b | 34774091261 |
+| #4033 | 84b197e76 | 34745744790 |
+
+#4309 is held out: it is 288 commits behind dev, and its provider-count
+assertions cannot be trusted until it is rebased.
+
+Two reviewer subagents read the diffs first. All five of batch 1's luvs01 pull
+requests came back MERGE with specific justification; the other four came back
+HOLD for exactly one reason each, the missing exact-head suite run, which is what
+the approvals above address.