diff --git a/CREDITS.md b/CREDITS.md
index 5de4f2f696..5737ed185e 100644
--- a/CREDITS.md
+++ b/CREDITS.md
@@ -243,6 +243,42 @@ unnecessary.
 If you find a landing that belongs on this page, open an issue. Being missed is
 the defect this file documents, not a claim you have to argue for.
 
+### 2026-09-13: independent work that overtook an open proposal
+
+The gate fires on what a pull request *says*. It cannot fire on a landing that
+never mentions the proposal it overtakes, which is how this one happened.
+
+[#4077](https://github.com/lidge-jun/opencodex/pull/4077) by
+[@laerad777](https://github.com/laerad777) proposed opening the xAI Grok OAuth
+lane to `service_tier: "priority"` and correcting the Fast-tier catalog copy.
+The registry half landed independently through #4431 at `7ca00ffe7`, derived
+from its own live probe, with no reference to #4077 and no trailer. The landed
+scope is narrower on evidence — `grok-4.20-multi-agent-0309` stays excluded
+because the gateway answers `service_tier: "default"` when sent `priority` —
+so this is genuinely independent work rather than a silent carry.
+
+The copy correction was still unlanded, and it was the part #4077 identified
+first. It landed through #4474 with a `Co-authored-by` trailer naming the
+author. The registry half is recorded here as an acknowledgement rather than as
+carried code, because that is what the evidence supports.
+
+The generalizable point: "independent" and "first" are different claims, and
+only the second one is visible from the open queue.
+
+### 2026-09-13: the gate also fires on prose about carrying
+
+The matcher reads the description, so a pull request that merely *describes* a
+carry train trips `missing_coauthor_credit` even when it has no source author.
+[#4499](https://github.com/lidge-jun/opencodex/pull/4499) was an ordinary
+implementation with no source branch; the phrases "contributor-carry train" and
+"Head commit carries `[skip ci]`" were enough to fail the gate. Rewording
+cleared it.
+
+That is a false positive rather than a defect worth loosening the matcher for.
+A gate that occasionally asks an author to justify wording is cheaper than one
+that misses a real uncredited carry, which is the failure this whole page
+documents. Write around it.
+
 ### A gap the gate does not close
 
 The gate checks that a trailer is **present**. It cannot check that the trailer
diff --git a/devlog/_plan/260913_contributor_carry_train/000_plan.md b/devlog/_plan/260913_contributor_carry_train/000_plan.md
new file mode 100644
index 0000000000..d3b03d7fe3
--- /dev/null
+++ b/devlog/_plan/260913_contributor_carry_train/000_plan.md
@@ -0,0 +1,137 @@
+# Contributor carry train — 60+ scored work into dev
+
+## Objective
+
+Land the highest-value contributor work that is still open on this repository.
+Seventeen open pull requests score 60 or higher by maintainer review once the
+maintainer's own #4462 is set aside; 16 of them are dispatched here and #3389 is
+deferred for a recorded reason. Eight issues score 60 or higher with no pull
+request owning them. The goal also closes everything those landings actually
+resolve. `origin/dev` was `2df82f412` when this roadmap was written.
+
+This unit follows the 36-PR lane-stack merge in
+`devlog/_plan/260913_lane_stack_merge/`, which landed the maintainer-authored
+backlog. That batch is the precedent for the mechanism here; what changes is the
+authorship. Every branch in this train belongs to someone else, so attribution is
+a correctness requirement rather than a courtesy.
+
+## Selection
+
+Candidates come from the maintainer's own `## 리뷰 · 우선순위 NN / 80` comments,
+harvested live from every open issue and pull request through the GraphQL API on
+2026-09-13. All 73 open pull requests and all 58 open issues carry a score, so
+the 60 cut is a real threshold rather than a sample.
+
+The inventory and its verification live in `001_candidate_inventory.md`.
+
+## Attribution is a gate, not a footnote
+
+`AGENTS.md` requires a `Co-authored-by` trailer naming the original author on any
+landing that reimplements, supersedes, carries, or rebases their pull request.
+`CREDITS.md` records 27 landings that failed this and explains why prose credit is
+not equivalent: GitHub reads the trailer, and nothing reads a sentence in a commit
+body.
+
+Two rules follow for every lane in this unit:
+
+- The trailer address is taken from the author's GitHub account, in the numeric
+  `users.noreply.github.com` form, not from the commit metadata on their branch.
+- The trailer is verified in the **actual landing commit** after the squash, not
+  in the pull request description. A custom squash message silently drops text
+  that was only in the body.
+
+## CI economy
+
+Unchanged from the previous batch and re-verified there. Each lane is a cumulative
+stack: the bottom branch merges `origin/dev`, each branch above merges its parent's
+resulting commit, so merging bottom-up produces shrinking diffs. Every non-tip head
+commit carries `[skip ci]`, which suppresses `ci.yml`, `react-doctor`,
+`service-lifecycle` and `issue-quality-tests`. Only `enforce-pr-target`,
+`pr-hygiene` and `pr-labeler` still run, because `pull_request_target` ignores the
+skip marker.
+
+Only the lane tip runs the full matrix, and that tip run is the merge gate for the
+whole lane. Squash messages never contain `[skip ci]`, because the dev-branch run
+each merge triggers is the regression gate.
+
+This is an owner-authorized deviation from the MAINTAINERS.md requirement that
+every pull request carry its own successful required check. Each merge comment
+states it explicitly: the owner authorization, the tip pull request and run id
+that covers the branch, and the fact that this branch's own `ci` check never ran.
+
+## Common principles
+
+Pushes use `git push --no-verify`, fast-forward only. No `--force` anywhere.
+Nothing is pushed to `dev`, `main` or `preview`.
+
+Local full-suite runs are forbidden. Allowed local checks are `bun run typecheck`,
+`bun run structure:check`, `bun run privacy:scan`, and `bun test` limited to the
+files a pull request touches. Hosted CI on the lane tip is the only suite proof
+this goal accepts.
+
+Each lane thread works in its own managed worktree and may fan out unlimited
+`xai/grok-4.6` subagents inside that worktree. Subagents share their parent's
+checkout, so a lane's subagents never run branch-level git operations concurrently.
+
+Lane threads never merge, never mark a pull request ready, and never close
+anything. They push and report. The main session performs every merge.
+
+Conflicts are resolved by reading both sides and judging which matches current
+behavior. A genuinely ambiguous conflict stops that link and is reported with both
+sides and the reasoning, never guessed past.
+
+## Lane map
+
+| Lane | Doc | Contents, bottom to top | Owner model |
+| --- | --- | --- | --- |
+| R responses/core | 010 | 4455, 4086, 4409, 4387 | anthropic/claude-opus-5 |
+| C chat + adapters | 010 | 4438, 4389, 4457 | anthropic/claude-opus-5 |
+| L cli + hub | 010 | 4382, 4413, 4170 | xai/grok-4.6 |
+| B bridge + images + auth | 010 | 4381, 4388, 4460 | xai/grok-4.6 |
+| S security-review hold | 010 | 4447 | xai/grok-4.6 |
+| X small carries | 010 | 4077 copy fix, 4171 dedupe | xai/grok-4.6 |
+| I1 Windows issues | 010 | 4425, 4442 | kimi/k3[1m] |
+| I2 config + account issues | 010 | 4430, 4435 | xai/grok-4.6 |
+| H context history | 030 | 3663 | anthropic/claude-opus-5 |
+| I3 routing compatibility | 030 | 3775, 4436, 4429 | xai/grok-4.6 |
+| I4 encrypted history regression | 030 | 4454 | anthropic/claude-opus-5 |
+
+Wave 1 prepares R, C, L, B, S, X, I1 and I2 in parallel. Wave 2 is I3, I4 and H.
+
+Wave 2 is not parallel throughout. I3 prepares alongside, but I4 and H must be
+serialized in that order: both land in the routed Responses request path, so H
+rebases onto I4 rather than preparing beside it. Preparing them as peers would
+put two `src/server/responses/core.ts` writers in one wave, which is the exact
+thing lane R exists to prevent.
+
+Lane S is prepared in wave 1 but is deliberately not merged in wp3. `#4447`
+touches CORS and the management provider routes, which is inside the
+MAINTAINERS.md security-review boundary, and a tip merge would have landed it on
+a CI signal that was never meant to certify it. Splitting it out of lane B keeps
+lane B's tip at `#4460` and keeps the hold visible at merge time rather than
+three documents away.
+
+## Work phases
+
+wp1 is this roadmap. wp2 prepares wave 1, wp3 merges it, wp4 prepares wave 2
+against landed `dev`, wp5 merges wave 2 and confirms the dev regression run, and
+wp6 closes what landed and records the outcome.
+
+## Risks
+
+Five of the carried pull requests touch `src/server/responses/core.ts`: #4455,
+#4086, #4409 and #4387 in lane R, plus #3663 in lane H. Issue #4454 lands in the
+same path without being a pull request at all. #3389 is a sixth `core.ts`
+toucher and is deferred for an unrelated reason recorded in
+`001_candidate_inventory.md`. Lane R serializes its four; H and I4 serialize
+after R. A lane merged out of order defeats the shrinking-diff property.
+
+GitHub had not computed `mergeable` for most of these branches when the roadmap
+was written, so lane assignment rests on file overlap rather than a proven
+conflict-free merge. Each lane thread discovers its real conflicts at prepare time
+and reports them.
+
+All four issue lanes — I1, I2, I3 and I4 — have no branch to carry at all. Those
+are ordinary implementations by the lane thread, and they carry no
+`Co-authored-by` trailer because there is no source branch; the reporter is
+credited in the description instead.
diff --git a/devlog/_plan/260913_contributor_carry_train/001_candidate_inventory.md b/devlog/_plan/260913_contributor_carry_train/001_candidate_inventory.md
new file mode 100644
index 0000000000..ce5c04a0f0
--- /dev/null
+++ b/devlog/_plan/260913_contributor_carry_train/001_candidate_inventory.md
@@ -0,0 +1,119 @@
+# Candidate inventory and its verification
+
+## How the scores were collected
+
+The maintainer leaves a `## 리뷰 · 우선순위 NN / 80` line in a comment on triaged
+issues and pull requests. This unit harvested that line from every open item on
+2026-09-13 through the GraphQL API.
+
+The first attempt was wrong in a way worth recording. `gh api graphql --paginate`
+only advances a query whose cursor variable is named `$endCursor`; a query using
+`$cursor` re-requests page one forever. That run produced a 230 MB file of 472
+identical pages and, worse, a plausible-looking result: deduplicating by number
+hid the loop and returned a correct-but-partial page-one answer. Five candidates
+at 61 to 68 were missing from it.
+
+The corrected collection covers 73 open pull requests and 58 open issues. Every
+one carries a score, so nothing in this inventory is an unscored guess.
+
+Seventeen open pull requests score 60 or higher once the maintainer's own #4462
+is set aside. Sixteen are dispatched here; #3389 is deferred for a reason
+recorded below.
+
+## Pull request candidates, 60 and above
+
+| Score | PR | Author | Domain | Lane |
+| --- | --- | --- | --- | --- |
+| 74 | #4457 | jeongjin0 | devin adapter tool identity | C |
+| 73 | #4381 | luvs01 | bridge truncated terminal | B |
+| 71 | #4387 | luvs01 | web-search continuation key binding | R |
+| 70 | #4455 | jeongjin0 | code-mode view_image | R |
+| 69 | #4388 | luvs01 | images loop buffering bound | B |
+| 68 | #3389 | Yum-wu | zero-output retry | deferred |
+| 68 | #4382 | luvs01 | hub status credentials | L |
+| 67 | #4438 | Yongzhaooo | OCG DeepSeek timeline instructions | C |
+| 66 | #4389 | olddonkey | stream memory allocations, 44 files | C |
+| 66 | #4413 | rrmlima | authenticated remote catalog pull | L |
+| 66 | #4460 | AgenticLab-SH | model availability vs auth failure | B |
+| 64 | #4077 | laerad777 | xAI Fast catalog copy, residue only | X |
+| 63 | #4086 | Eleven-is-cool | routed continuation after replay miss | R |
+| 61 | #4409 | yxr1995-maker | routed effort ladders from models.dev | R |
+| 61 | #3663 | y2ambition-ai | context history relay, 19 files | H |
+| 61 | #4170 | yeongjunyoo | stop refusal cause | L |
+| 61 | #4447 | Veritas-7 | field-masked provider writes | S |
+
+`#4462` also scores 74 but is maintainer-authored and belongs to the sub-agent
+surface work, not this train.
+
+`#3389` is deferred by prior judgment, not by score. The 2026-09-04 priority-65
+closeout recorded `NEEDS_DESIGN`: its zero-output premise was disproved by
+experiment. Reviving it needs a design decision first, so it is not dispatched
+here.
+
+## Issue candidates with no owning pull request
+
+| Score | Issue | Reporter | Lane |
+| --- | --- | --- | --- |
+| 76 | #4425 | wanjinxingoo-bot | I1 |
+| 76 | #4430 | samwang0041-star | I2 |
+| 73 | #4454 | 321sssrt-bit | I4 |
+| 71 | #3775 | leonclab | I3 |
+| 70 | #4429 | mdwsk88 | I3 |
+| 67 | #4442 | SeanChengN | I1 |
+| 66 | #4436 | jaychou0642-create | I3 |
+| 64 | #4435 | ren-min-wan-sui | I2 |
+
+The remaining 60+ issues are excluded for a stated reason rather than by
+oversight. #4456, #4412 and #4439 already have an owning pull request in this
+train (#4457, #4455, #4438). #4191, #4311, #3661, #3781, #4312 and #3506 are
+partially landed: the merged work references them with `Refs` and each landing
+states in its own description that the issue stays open. #3375, #3376 and #3377
+are maintainer feature issues under separate execution.
+
+## Supersede closures already executed
+
+Before this train was planned, nine contributor pull requests were closed because
+the previous batch had already absorbed them. Each carrier's merge commit was
+verified as an ancestor of `origin/dev` at `2df82f412` before the close, and each
+close carries a comment naming the carrier, the merge commit, and the fact that
+the carry preserved the author's trailer.
+
+| Closed | Author | Carrier | Merge commit |
+| --- | --- | --- | --- |
+| #4319 | ke-1t | #4346 | `b551b524f` |
+| #4310 | yxr1995-maker | #4354 | `b474013` |
+| #3652 | itismyfield | #4355 | `c6372ca` |
+| #4229 | yansigit | #4363 | `37bc1a0` |
+| #4216 | ildunari | #4367 | `c240534` |
+| #4119 | DamnUi | #4366 | `90667e5` |
+| #4317 | cortes-ventures | #4402 | `8acd73b` |
+| #4080 | terrytan95 | #4369 | `c27eeec` |
+| #3458 | Ingwannu | #4362, #4372 | `19601ea`, `7874900` |
+
+## The half-superseded case
+
+`#4077` was not closed, and the reason generalizes. Its registry half landed
+independently through `#4431` (`7ca00ffe7`), which classified the xAI OAuth lane
+from its own live probe and deliberately excluded `grok-4.20-multi-agent-0309`
+because the gateway answers `service_tier: "default"` when sent `priority`. That
+branch flips the same model to `true`, so the landed scope is narrower on
+evidence.
+
+Its third point is still unlanded and now more visible than when it was filed:
+`src/providers/registry.ts` still reads
+`fastTierDescription: "Priority processing, 2x token price"`, and `#4431` opened
+the OAuth subscription rows where no per-token price exists. Lane X carries that
+correction with the author's trailer.
+
+`#4431` landed without referencing `#4077` and without a trailer for its author.
+That is the `missing_coauthor_credit` pattern `CREDITS.md` exists to stop
+repeating, and it is why lane X exists as its own slice rather than as a footnote
+in another lane.
+
+## Duplicate to collapse
+
+`#4171` (rrmlima, `CHANGES_REQUESTED`) and `#4455` (jeongjin0) both answer issue
+`#4412` and both touch `src/responses/code-mode-helper-compat.ts` and
+`src/types/tools.ts`. `#4455` is the 8-file superset. Lane R carries `#4455` with
+trailers for both authors; lane X closes `#4171` after that landing is verified on
+`dev`.
diff --git a/devlog/_plan/260913_contributor_carry_train/010_wave1.md b/devlog/_plan/260913_contributor_carry_train/010_wave1.md
new file mode 100644
index 0000000000..c016253407
--- /dev/null
+++ b/devlog/_plan/260913_contributor_carry_train/010_wave1.md
@@ -0,0 +1,205 @@
+# wp2 — Wave 1 lane preparation
+
+Eight worktree threads prepare in parallel. Every head in this train lives on a
+contributor fork, so no lane pushes to a source branch. Each lane creates its own
+`codex/260913-carry-*` branches in `lidge-jun/opencodex` and opens new pull
+requests that carry the work with attribution.
+
+## The carry shape, identical for every link
+
+1. Branch from `origin/dev` for the bottom link, or from the previous link's
+   pushed head for every link above it.
+2. Fetch the source head directly: `git fetch https://github.com/<owner>/opencodex <branch>`.
+   The fork is not a configured remote in a fresh worktree.
+3. Apply the source content. Cherry-pick when it is clean, reimplement when the
+   branch is stale against current `dev`. Either way the result is judged against
+   current behavior, not against the branch's original base.
+4. Fold in the review findings already on the source pull request. A
+   `CHANGES_REQUESTED` review is not an approval and its findings are not optional.
+5. Add `Co-authored-by: <Name> <id+login@users.noreply.github.com>` to the commit
+   that will survive the squash, taking the address from the author's GitHub
+   account rather than from their commit metadata.
+6. Non-tip head commit subjects carry `[skip ci]`. The tip does not.
+7. Push with `git push --no-verify`, fast-forward only.
+8. Open the pull request against `dev` for the bottom link and against the parent
+   link's branch for every link above, filling every section of
+   `.github/PULL_REQUEST_TEMPLATE.md`.
+
+A lane reports: each new pull request number, the tip head SHA, the tip CI run id
+and its conclusion, and every conflict it resolved with the reasoning.
+
+## Lane R — responses/core, serialized
+
+Owner `anthropic/claude-opus-5`. Order is fixed because all four touch
+`src/server/responses/core.ts`.
+
+| # | Source | Author | Fork branch | Head |
+| --- | --- | --- | --- | --- |
+| 1 | #4455 | jeongjin0 | `codex/code-mode-view-image-helper` | `61b9aefc3` |
+| 2 | #4086 | Eleven-is-cool | `fix/routed-continuation-replay` | `0c39bc608` |
+| 3 | #4409 | yxr1995-maker | `feat/reasoning-metadata` | `7de03fb06` |
+| 4 | #4387 | luvs01 | `agent/web-search-key-binding-20260912` | `2a1b3c307` |
+
+`#4455` carries two trailers. It answers issue #4412 together with `#4171`
+(rrmlima), whose four files are a strict subset of its eight, so both authors are
+named. Lane X closes `#4171` after this lands.
+
+`#4409` adds `scripts/test-layout/layout.json` and
+`tests/fixtures/test-layout-expected.json` entries. `tests/test-layout.test.ts`
+and `tests/test-layout-tooling.test.ts` enforce both; a new test file missing
+from either is a hard failure, not a lint warning.
+
+`#4387` touches fifteen `structure/` documents. `bun run structure:check` is a
+required local check for this link.
+
+## Lane C — chat and adapters
+
+Owner `anthropic/claude-opus-5`. `#4438` and `#4389` share
+`src/adapters/openai-chat.ts`, so the order is fixed.
+
+| # | Source | Author | Fork branch | Head |
+| --- | --- | --- | --- | --- |
+| 1 | #4438 | Yongzhaooo | `codex/ocg-deepseek-system-order` | `9ebbcad26` |
+| 2 | #4389 | olddonkey | `codex/memory-stream-optimizations` | `176cbbd2e` |
+| 3 | #4457 | jeongjin0 | `codex/devin-restore-tool-names` | `edc6db3ff` |
+
+`#4389` is the cost of this lane: 44 files, +1491/-268, touching four adapters,
+the chat outbound path, admission, the translator budget and request
+decompression. It is a performance change, so the audit question is whether any
+allocation removal changes observable behavior on a truncated or folded stream.
+
+`#4438` closes issue #4439 and `#4457` closes issue #4456 once landed; both
+closures belong to wp6, not to the lane.
+
+## Lane L — cli and hub
+
+Owner `xai/grok-4.6`. Three links with no shared file at all. They are chained
+only to collapse three CI runs into one; no ordering dependency exists between
+them.
+
+| # | Source | Author | Fork branch | Head |
+| --- | --- | --- | --- | --- |
+| 1 | #4382 | luvs01 | `agent/status-hub-binding-20260912` | `094f12713` |
+| 2 | #4413 | rrmlima | `feat/remote-catalog-pull` | `2aa821dff` |
+| 3 | #4170 | yeongjunyoo | `fix/4169-stop-refusal-cause` | `4d72ef010` |
+
+`#4413` adds a CLI command, so `bun run skill:surface` must be regenerated and
+`bun run skill:surface:check` must pass — `tests/ci-workflows/skill-ocx.test.ts`
+fails on a committed map that drifts from `src/cli/capabilities.ts`, and it also
+fails on a documented command the registry does not have.
+
+`#4413` duplicates its author's own issue #3729. That issue closes with this
+landing.
+
+## Lane B — bridge, images, auth
+
+Owner `xai/grok-4.6`.
+
+| # | Source | Author | Fork branch | Head |
+| --- | --- | --- | --- | --- |
+| 1 | #4381 | luvs01 | `agent/content-filter-terminal-20260912` | `aad1d75bc` |
+| 2 | #4388 | luvs01 | `agent/media-loop-bounds-20260912` | `9518b5281` |
+| 3 | #4460 | AgenticLab-SH | `codex/260913-model-capacity-errors-dev` | `eefa07442` |
+
+`#4460` ships a `devlog/_fin/` document on its branch. A `_fin` unit records work
+already visible in public git history, so the carry either moves it to `_plan` or
+drops it; it does not land a `_fin` record for work that has not landed.
+
+The lane tip is `#4460`.
+
+## Lane S — security-review hold
+
+Owner `xai/grok-4.6`. One link: `#4447` (Veritas-7, fork branch
+`fix/openai-patch-operator-overlays`, head `ed9655286`), letting field-masked
+writes reach canonical OpenAI past stored overlays.
+
+It is its own lane rather than the top of lane B for a reason that is easy to get
+wrong. It touches `src/server/auth-cors.ts` and
+`src/server/management/provider-routes.ts`, which is inside the MAINTAINERS.md
+security-review boundary. If it sat on top of lane B, merging that lane's tip in
+wp3 would land the CORS and management change as a side effect of a lane merge,
+on a CI signal that was never meant to certify it.
+
+`#4447` targets `main`, which is why its own title says `[WRONG BRANCH]`. The
+carry lands on `dev` like everything else; the source pull request is left for
+wp6 disposition rather than retargeted underneath its author.
+
+This lane prepares in wave 1 and does not merge in wp3. It merges in wp5 only
+after maintainer security review, or it is recorded as deferred with that reason.
+
+## Lane X — small carries and dedupe
+
+Owner `xai/grok-4.6`. One branch, two obligations.
+
+The carry is the residue of `#4077` (laerad777, fork branch
+`fix/xai-oauth-service-tier`, head `a071c67a1`, currently `CONFLICTING`): change
+`fastTierDescription` on the xai registry entry away from
+`"Priority processing, 2x token price"`, because `#4431` opened OAuth subscription
+rows that have no per-token price. The registry classification in that branch is
+already landed and is **not** re-applied; in particular
+`grok-4.20-multi-agent-0309` keeps `forwardCallerServiceTier: false`, since a live
+probe recorded the gateway downgrading it to `default`.
+
+The pins that read this string must move with it. The lane finds them rather than
+assuming: `rg fastTierDescription` across `src/`, `tests/` and `docs-site/`.
+
+The dedupe is `#4171`, which wp6 closes after lane R's `#4455` carry is verified
+on `dev`. Lane X does not close it.
+
+## Lane I1 — Windows issues, no source branch
+
+Owner `kimi/k3[1m]`. Two implementations, no carry, so no `Co-authored-by`; the
+reporters are credited in the descriptions and the issues are linked.
+
+`#4425` (76): a non-elevated `ocx service install` fails because the generated
+`LogonTrigger` omits `UserId`, and the resulting `Access is denied` is reported as
+a permission problem, which is a misdiagnosis. The fix emits `UserId` and the
+diagnostic stops blaming elevation.
+
+`#4442` (67): the Windows history manifest integrity check treats `C:\...` and
+`\\?\C:\...` as different paths, so the check fails after sessions are archived.
+The fix normalizes the extended-length prefix before comparison.
+
+Both are Windows-only paths that this lane cannot execute locally, and the obvious
+evidence is not available by default: `platform-windows` in
+`.github/workflows/ci.yml` is gated on `github.event_name == 'workflow_dispatch'`,
+so a pull-request tip run never executes it. The gate for this lane is therefore
+an explicit `workflow_dispatch` run on the exact tip SHA, plus focused tests that
+encode the path and trigger shapes directly. If the dispatch run is not obtainable,
+the lane says so instead of implying a Windows leg ran.
+
+## Lane I2 — configuration and account issues, no source branch
+
+Owner `xai/grok-4.6`.
+
+`#4430` (76): 2.52.0 treats `claudeCode.desktopProfile.appliedFingerprint=null` as
+fatal and replaces the entire configuration with defaults. Losing a user's whole
+config on a nullable field is the severe part; the fix accepts `null` and, more
+generally, stops a single invalid field from triggering a whole-config reset.
+
+`#4435` (64): a second Kiro account disappears shortly after a successful login.
+The lane reproduces against the account store before proposing a fix, because a
+disappearing second account is equally consistent with a write collision, a
+single-slot overwrite, and a projection that keys on a non-unique identifier.
+
+## Dispatch packet
+
+Every thread receives the same frame, with its lane table substituted:
+
+- Worktree environment, this repository, base `dev`.
+- The common principles from `000_plan.md`, quoted rather than referenced.
+- Unlimited `xai/grok-4.6` subagents may be fanned out inside the lane's own
+  worktree, in parallel. They share that worktree's checkout, so they need
+  non-overlapping write scopes and must never run branch-level git operations
+  concurrently.
+- Do not merge, do not mark ready, do not close anything, do not push to `dev`,
+  `main` or `preview`.
+- Report the pull request numbers, the tip head SHA, the tip run id and
+  conclusion, and every conflict resolution with its reasoning.
+
+## Exit criteria
+
+Every wave-1 lane has its chain pushed, its pull requests open with complete
+descriptions, and a hosted run on its exact tip head. A lane whose tip run is red
+for a reason belonging to the lane fixes it before wp3; a lane whose tip is red
+because `dev` drifted re-merges `origin/dev` at its tip and re-runs.
diff --git a/devlog/_plan/260913_contributor_carry_train/011_wave1_outcome.md b/devlog/_plan/260913_contributor_carry_train/011_wave1_outcome.md
new file mode 100644
index 0000000000..a073b351ff
--- /dev/null
+++ b/devlog/_plan/260913_contributor_carry_train/011_wave1_outcome.md
@@ -0,0 +1,116 @@
+# wp2/wp3 — Wave 1 outcome
+
+All eight wave-1 lanes are on dev. Lane S landed last, after the security review
+it was held for changed the diff.
+
+## What landed
+
+| Lane | Tip | Merge commit | Carried |
+| --- | --- | --- | --- |
+| L | #4481 | d865aacf93 | #4382 luvs01, #4413 rrmlima |
+| B | #4480 | 2af30c2d0e | #4381 luvs01, #4388 luvs01, #4460 AgenticLab-SH |
+| X | #4474 | 2296e485d6 | #4077 laerad777, residue only |
+| I1 | #4486 | a3ca64f605 | issues #4425, #4442 |
+| I2 | #4482 | 990cd8cce5 | issues #4430, #4435 |
+| C | #4487 | 55bb9f3fef | #4438 Yongzhaooo, #4389 olddonkey, #4457 jeongjin0 |
+| R | #4489 | 3f76ce415d | #4455 jeongjin0, #4409 yxr1995-maker, #4387 luvs01 |
+| S | #4477 | 981b53e7d0 | #4447 Veritas-7, plus the review fix c39098ba3d |
+
+Every merge was gated the same way: a Cross-platform CI run concluded success on
+the exact tip head SHA, the merge commit was verified with
+git merge-base --is-ancestor against origin/dev afterwards, and the
+Co-authored-by trailers were read out of the landed commits rather than the pull
+request bodies.
+
+## The security hold earned its keep
+
+#4477 was green on df7cbd5b7b for hours before it merged. It carries #4447, which
+touches src/server/auth-cors.ts and src/server/management/provider-routes.ts, so
+MAINTAINERS.md requires explicit security review and a green tip is not that
+review. Splitting it out of lane B is what made the hold enforceable: as lane B's
+tip it would have landed as a side effect of a lane merge.
+
+The review changed the outcome, which is the argument for the hold existing at
+all. The threat model established that overlay tolerance is openai-only, that the
+destination and auth keys stay byte-pinned, and that PATCH is an explicit
+per-field allowlist so a request cannot introduce a novel key. It also established
+that the canonical OpenAI seed defines only four keys — adapter, authMode, baseUrl,
+codexAccountMode — so "ignore keys the seed never defines" reaches nearly every
+config key, which is a much wider door than the description implied.
+
+That width is where the finding was. The author had already denied
+allowPrivateNetwork, correctly: it is patchable, it disables destination DNS
+classification, and overlay tolerance would have persisted it on the ChatGPT
+forward row. headers sits in exactly the same class and was not covered. Canonical
+OpenAI has no registry staticHeaders, the PATCH field mask writes headers with a
+shallow merge, and the forward adapter applies provider.headers to the upstream
+request before the incoming forward headers — so a persisted value wins whenever
+the caller omits that header. A dashboard-session
+PATCH {"headers":{"chatgpt-account-id":"..."}} would have ridden every subsequent
+ChatGPT request that did not carry the header itself. POST still refused it; PATCH,
+the editor and reload did not.
+
+c39098ba3d denies headers on canonical openai the same way and adds the missing
+PATCH regression. The test was driven red before it was accepted: removing the
+guard fails exactly that case and nothing else in the file.
+
+Two process notes worth keeping. The finding came from an independent reviewer
+rather than the main pass, which had stopped at the field-policy map and concluded
+headers were redacted — true for editor admission, false for the PATCH mask, which
+has its own allowlist. And the tip's first run failed in select-windows-runner with
+no failing step; re-running the failed jobs on the same commit turned it green, so
+the exact-head evidence survived rather than needing a new head.
+
+Recorded follow-up: the overlay tolerance is a denylist and denylists rot.
+PROVIDER_CONFIG_FIELD_POLICY forces a new provider field to be classified but does
+not force an overlay decision, so a future editor field touching a trust boundary
+becomes silently reachable on the canonical row. codexToolMode is the current
+example. The durable fix is an explicit overlay allowlist plus a guard test.
+
+## The credit defect this wave surfaced
+
+Lane I1 implemented #4442 as an ordinary fix, because the candidate harvest found
+no pull request owning that issue. There was one: draft #4465 by maoxin1234,
+opened 2026-09-13T05:43Z, after the harvest and before the lane. It proposed the
+same normalization in the same two files.
+
+The landed implementation is a superset — it adds the backup-id compatibility
+fallback, the native-residue path and the structure doc — but it supersedes a
+contributor proposal that came first, which AGENTS.md treats as requiring a
+trailer. The I1 merge commit therefore names maoxin1234, and #4465 was closed with
+that stated rather than closed as a duplicate.
+
+The generalizable failure is the harvest, not the lane. A scored inventory is a
+snapshot, and an issue lane must re-check for an owning pull request at dispatch
+time rather than trusting the snapshot it was planned from. Wave 2 lanes I3 and I4
+do that check before implementing.
+
+## What #4086 turned out to be
+
+Lane R's fourth planned link was already on dev as d6723f7f3, with its own
+Co-authored-by trailer for Eleven-is-cool. The lane attempted the carry before
+concluding that, and the modify/delete conflict on
+structure/04_transports-and-sidecars.md — a file the #4276 SSOT restructure had
+deleted — is what prompted the check. The landed version is a superset of the
+branch.
+
+That is the second time in this train that a planned carry was already satisfied
+on dev; #4170 in lane L was the first. Both were found by attempting the work
+rather than by reading the plan, which is the argument for lanes re-verifying
+their own inputs.
+
+## Disposition executed
+
+Closed with evidence comments naming the landing pull request and merge commit:
+source pull requests #4382, #4413, #4170, #4381, #4388, #4460, #4077, #4438,
+#4389, #4457, #4455, #4409, #4387, #4171, #4086, #4465, plus the carry links that
+GitHub did not auto-close (#4479, #4473, #4485). Issues #4425, #4442, #4430,
+#4435, #4439, #4456, #4412 and #3729 were closed the same way.
+
+## Honest limits
+
+Non-tip pull requests merged without their own ci check, under the recorded owner
+authorization for tip-only CI. Every dev run triggered by these merges ended
+cancelled by the concurrency group as the next merge superseded it, which is the
+workflow behaving as configured; the batch regression gate is a completed dev run
+on a commit containing everything, and that belongs to wp5 rather than here.
diff --git a/devlog/_plan/260913_contributor_carry_train/020_wave1_merge.md b/devlog/_plan/260913_contributor_carry_train/020_wave1_merge.md
new file mode 100644
index 0000000000..7b0ea43350
--- /dev/null
+++ b/devlog/_plan/260913_contributor_carry_train/020_wave1_merge.md
@@ -0,0 +1,71 @@
+# wp3 — Wave 1 integration
+
+The main session performs every merge. No lane thread merges anything.
+
+## Order
+
+Lanes merge one at a time, smallest blast radius first, so that each later lane
+re-merges a dev that already contains the earlier ones:
+
+X, I1, I2, L, B, C, R. Lane S does not merge here.
+
+X is first because it is a single-file registry correction. R is last because its
+four links all rewrite src/server/responses/core.ts, which is the file every other
+lane is most likely to have touched indirectly through
+structure/transports/responses.md.
+
+## The wave-1 hold
+
+Lane S (#4447) prepares in wave 1 and is deliberately excluded from this merge
+order. It touches CORS and the management provider routes, which is inside the
+MAINTAINERS.md security-review boundary, so it merges in wp5 after review or is
+recorded as deferred. Lane B's tip is #4460; if a merge attempt on lane B would
+carry #4447, the lane was assembled wrong and is fixed before merging rather than
+merged and reverted.
+
+## Per-lane merge procedure
+
+1. git fetch origin dev and re-read the head. dev moves during this batch.
+2. Read the lane tip current head SHA with gh pr view <tip> --json headRefOid.
+3. Confirm the hosted run that concluded success ran on that exact SHA. A green
+   run on an earlier head is not evidence for the head being merged.
+4. If the tip re-merged origin/dev after its green run, verify the resolution by
+   reading it and re-run bun run typecheck, bun run structure:check and
+   bun run privacy:scan before merging. This happened twice in the previous batch
+   and both times the re-merge was mechanical; that is a finding to reconfirm,
+   not to assume.
+5. Merge the tip. A cumulative lane merges as one merge commit on the tip, which
+   lands every link beneath it. A lane whose links must appear separately in
+   history squash-merges bottom-up instead.
+6. Verify the landing rather than trusting the merge report:
+   git merge-base --is-ancestor <merge-sha> origin/dev.
+7. Verify attribution in the landed commit: git log -1 --format=%B <merge-sha>
+   must show the Co-authored-by trailers for every carried author. A trailer that
+   lived only in the pull request body is gone after a custom squash message, and
+   that is exactly the failure CREDITS.md documents.
+
+## Merge comment
+
+Every merge comment names three things explicitly, so the deviation stays a
+recorded owner decision rather than an inferred one:
+
+- the owner authorization for tip-only CI in this batch,
+- the tip pull request and run id that covers this branch,
+- the fact that this branch own ci check never ran.
+
+## The screenshot gate
+
+enforce-target fails with missing UI screenshot on any pull request whose title or
+description merely mentions gui. No wave-1 source pull request touches gui/, so
+the gate is expected to be quiet here — but it reads the text, not the diff, so a
+description that mentions the dashboard trips it anyway. Where a real rendered
+state exists, capture it. A control that only renders against a real expired
+credential is covered by its tests, and that limitation is stated in the
+description rather than faked with a fixture.
+
+## Exit criteria
+
+Every wave-1 lane except S reports MERGED with its own merge commit, every merge
+commit is a verified ancestor of origin/dev, and every carried author trailer is
+present in the landed commit. Lane S exits wp3 prepared, green and unmerged; that
+is its success state here, not a failure to land.
diff --git a/devlog/_plan/260913_contributor_carry_train/030_wave2.md b/devlog/_plan/260913_contributor_carry_train/030_wave2.md
new file mode 100644
index 0000000000..7b3625a1b2
--- /dev/null
+++ b/devlog/_plan/260913_contributor_carry_train/030_wave2.md
@@ -0,0 +1,126 @@
+# wp4 — Wave 2 lane preparation
+
+Three lanes were held out of wave 1 because each collides with wave-1 work on a
+file wave 1 rewrites. They prepare against landed dev, not against the heads they
+would have started from.
+
+They are not all parallel. I3 prepares on its own. I4 and H both land in the
+routed Responses request path, so they run in that order and H rebases onto I4.
+Preparing them as peers would put two writers on src/server/responses/core.ts in
+one wave, which is precisely what lane R exists to prevent.
+
+## Lane H — context history relay: NOOP, already on dev
+
+Owner anthropic/claude-opus-5. Source #3663 (y2ambition-ai, fork branch
+feat/codex-context-history, head 8e0b53b0f), 19 files, +1249/-15.
+
+Outcome: there was nothing to carry. The lane verified against landed dev before
+writing anything and found the content already there as a33b51eb, landed through
+#4360 on 2026-09-12 (merge 2526715fb2), carried from exactly this branch's head
+with Co-authored-by trailers for both y2ambition-ai and nbbb26 intact. Five
+hardening commits followed it; all 19 files are present on dev and every one is
+larger there, so re-applying the branch would have been a downgrade.
+
+The lane also disproved the conflict this document predicted, and the reason is
+stronger than "different regions". Lane I4's strip is gated on
+adapter === "openai-responses" and not isCanonicalOpenAiForwardProvider, while
+the context-history ownership recording is gated on isCanonicalOpenAiForwardProvider
+and contextRelayActivated(). The predicates are complements on the same
+destination question, so no request takes both paths, and they sit on opposite
+sides of dispatch: one repairs the outbound body before the send, the other
+records the account after upstream accepted it.
+
+This is the third planned carry in this train that turned out to be already
+satisfied on dev, after #4170 in lane L and #4086 in lane R. All three were found
+by attempting the work rather than by reading the plan. #3663 was closed with a
+comment naming a33b51eb and #4360.
+
+Because lane H needed nothing, lane I5 branches from dev directly rather than
+from lane H's head.
+
+The original deferral reasoning is kept below, because it is why the lane was
+scheduled here at all.
+
+It is deferred because it touches src/server/responses/core.ts,
+src/server/index.ts and src/server/live.ts at once. Lane R rewrites the first and
+lane B touches the third, so preparing it in parallel would mean resolving the
+same conflicts twice.
+
+src/server/index.ts carries a hard constraint that a large carry can break
+silently. It is the composition root, and the window between the Bun.serve call
+and the labActivationRequired check must stay synchronous: a scan in that file
+fails on any await added to that window and on startServer being declared async.
+The reason is not style — the synchronous subagent-fallback chain has nowhere to
+await, so an await there reroutes subagents to a different model than the
+operator configured.
+
+The lane also adds test-layout entries and must keep
+tests/lab/core-lab-boundary.test.ts green: src/server/responses/core.ts may not
+reach src/lab/ even transitively.
+
+## Lane I3 — routing compatibility, no source branch
+
+Owner xai/grok-4.6. Three issues in the model and route capability area, deferred
+because #3775 overlaps lane R #4409 effort-ladder work.
+
+#3775 (71): Codex 0.153.4 rejects minimal and none on custom models mapped to
+gpt-6-astra. #4349 landed effort ceilings enforced independently of model pins, so
+this lane starts by determining what that already fixed; the honest outcome may be
+a narrower fix than the issue describes.
+
+#4436 (66): deepseek-flash accepts native image input on both Chat and Responses,
+but the registry still routes it through the vision sidecar. This one is a CARRY,
+not an implementation. The dispatch-time re-check found #4467 by
+jaychou0642-create, opened 2026-09-13T06:06Z — after the candidate harvest and
+before this lane — by the same person who filed the issue. It is ready rather than
+draft, targets dev, and touches 13 files (+154/-15) centred on
+src/providers/registry.ts, with locale docs and parity tests.
+
+Carry it with a Co-authored-by trailer naming
+jaychou0642-create <283093853+jaychou0642-create@users.noreply.github.com>. The
+review question is whether it declares the capability through the per-model
+contract #4374 and #4376 landed, rather than adding a new special case.
+
+That this was caught at all is the wave-1 lesson applied: the scored inventory is
+a snapshot, and lane I1 landed #4442 as an ordinary fix while contributor draft
+#4465 already proposed it. Every issue lane re-checks for an owning pull request
+at dispatch time now.
+
+#4429 (70): a key-auth Responses gateway (Kimi K3) echoes hosted web_search as a
+client function_call, and webSearchBridge only has an Ollama executor. This is the
+largest of the three and may split into its own work-phase if the executor
+abstraction turns out to be Ollama-shaped rather than merely Ollama-only.
+
+## Lane I4 — encrypted history regression
+
+Owner anthropic/claude-opus-5. #4454 (73) is an issue, not a pull request: there
+is no branch to carry and no Co-authored-by trailer. Reported by 321sssrt-bit,
+mixed encrypted agent_message history bypasses the routed Responses fail-closed
+path.
+
+Deferred behind lane R because it lands in the same request path, and it is a
+regression in a fail-closed boundary, which makes it the highest-risk single item
+in the train. A bypass of fail-closed is a security-boundary defect: it reports
+for maintainer review per MAINTAINERS.md rather than being treated as ordinary
+once CI is green. Lane H starts from this lane's pushed head, not from dev.
+
+#4364 bounded multipart encrypted task recovery and #4351 restored opt-in
+plaintext V2 messages; both explicitly state they do not close #3661. This lane
+establishes whether #4454 is a distinct defect or the same envelope problem seen
+from the other side before writing a fix.
+
+## Exit criteria
+
+Each wave-2 lane has its chain pushed against landed dev, its pull requests open,
+and a hosted run concluded on its exact tip head. I4 reaches that state before H
+starts.
+
+## Conditional closures
+
+Two of these lanes may find that the reported defect is already fixed or is a
+restatement of another open issue. #3775 may be covered by the effort ceilings
+#4349 landed, and #4454 may be the same envelope problem as #3661 seen from the
+other side. In either case the honest outcome is a comment on the issue naming
+what was found, not a closure claimed against a fix that was not written here.
+050_disposition.md lists the expected closures; a lane finding no defect
+overrides that list.
diff --git a/devlog/_plan/260913_contributor_carry_train/040_wave2_merge_regression.md b/devlog/_plan/260913_contributor_carry_train/040_wave2_merge_regression.md
new file mode 100644
index 0000000000..fbbcfc1597
--- /dev/null
+++ b/devlog/_plan/260913_contributor_carry_train/040_wave2_merge_regression.md
@@ -0,0 +1,46 @@
+# wp5 — Wave 2 integration and the dev regression gate
+
+## Order
+
+S, I3, I4, H. Lane S is first because it has been waiting since wave 1 and is a
+single link. H is last because it is the largest diff and the one most exposed to
+everything that landed before it, and because it was prepared on top of I4.
+
+The merge procedure is identical to 020_wave1_merge.md, including the exact-head
+CI check, the ancestry verification, the trailer check in the landed commit, and
+the three-part merge comment.
+
+## Security-review holds
+
+Two lanes do not merge on green CI alone. Lane S (#4447) touches CORS and the
+management provider routes; lane I4 (#4454) repairs a fail-closed bypass. Both are
+inside the MAINTAINERS.md security-review boundary. If review is not available
+within this goal, they are recorded as deferred with the reason rather than merged
+on a CI signal that was never meant to certify them.
+
+Deferring lane I4 has a consequence worth stating up front: lane H was prepared on
+top of it. If I4 is held, H is rebased onto dev without it before merging, or H is
+held too. H is not merged with an unmerged parent silently folded in.
+
+## The regression gate
+
+A merge report is not a regression proof. The gate is a dev workflow run that
+concluded success on a commit that contains the whole batch.
+
+Two distinctions carried over from the previous batch, both learned the hard way:
+
+- A run that ends cancelled is not a pass. Merges inside one concurrency group
+  supersede each other, which is the workflow behaving as configured; the evidence
+  is the completed run on a descendant commit.
+- The commit the run executed on must be an ancestor-verified descendant of the
+  last merge. git merge-base --is-ancestor <last-merge> <run-head> is the check.
+
+If the final run is red, the failure is triaged before any completion claim. A
+failure that belongs to a landed lane is fixed as a follow-up pull request in this
+same goal, not recorded as an acceptable residue.
+
+## Exit criteria
+
+Every lane is merged or explicitly deferred with a stated reason, and one dev run
+has concluded success on a batch-containing commit, recorded by run id and head
+SHA.
diff --git a/devlog/_plan/260913_contributor_carry_train/050_disposition.md b/devlog/_plan/260913_contributor_carry_train/050_disposition.md
new file mode 100644
index 0000000000..e447d56506
--- /dev/null
+++ b/devlog/_plan/260913_contributor_carry_train/050_disposition.md
@@ -0,0 +1,73 @@
+# wp6 — Disposition and outcome
+
+## What closes, and on what evidence
+
+A merged pull request does not prove its linked issue is resolved. The rule from
+the previous batch holds: Refs #N is not Closes #N, and a description stating the
+issue stays open outranks any topical similarity.
+
+GitHub auto-closes linked issues only on merge into the default branch. These
+merge into dev, so every closure here is manual.
+
+### Source pull requests carried by this train
+
+Each source is closed only after its carry is verified on dev by
+git merge-base --is-ancestor, with a comment naming the carry, its merge commit,
+and the trailer that preserved the author credit.
+
+| Source | Author | Carried by lane |
+| --- | --- | --- |
+| #4455, #4086, #4409, #4387 | jeongjin0, Eleven-is-cool, yxr1995-maker, luvs01 | R |
+| #4438, #4389, #4457 | Yongzhaooo, olddonkey, jeongjin0 | C |
+| #4382, #4413, #4170 | luvs01, rrmlima, yeongjunyoo | L |
+| #4381, #4388, #4460 | luvs01, luvs01, AgenticLab-SH | B |
+| #4447 | Veritas-7 | S |
+| #4077 | laerad777 | X |
+| #3663 | y2ambition-ai | H — no carry needed; already on dev as a33b51eb via #4360 |
+
+#4171 (rrmlima) closes against lane R #4455 carry as a duplicate, with both
+authors named in that landing.
+
+### Issues expected to close, and three that are conditional
+
+#4412 with lane R, #4439 and #4456 with lane C, #3729 with lane L, #4425 and
+#4442 with lane I1, #4430 and #4435 with lane I2, and #4429 with lane I3.
+
+#3775 and #4436 in lane I3, and #4454 in lane I4, are conditional rather than
+expected, for the reasons below.
+
+Each closure is re-verified against landed dev before it is executed, because the
+map above is written while the lanes are still in flight.
+
+Three of those closures are conditional rather than expected. #3775 may already be
+covered by the effort ceilings #4349 landed, #4436 may reduce to a capability
+declaration on the contract #4374 and #4376 established, and #4454 may be the same
+envelope defect as #3661 seen from the other side. If a lane finds no defect to
+fix, the issue gets a comment naming what was found and stays open. A closure is
+never claimed against a fix that was not written.
+
+### Issues that stay open
+
+#4191, #4311, #3661, #3781, #4312 and #3506 are partially landed. They receive a
+comment naming what landed and what remains, and they stay open. The previous
+batch checked 24 closure candidates and every one came back KEEP for exactly this
+reason: related work had shipped, but the actual ask had not been met.
+
+## Credit repair
+
+#4431 landed the xAI OAuth Fast classification without a trailer for the author of
+#4077, who had proposed the registry change first. The landing was independently
+derived from a live probe, so this is not a silent carry, but CREDITS.md is the
+place that distinction gets recorded rather than left to memory. wp6 adds the
+entry.
+
+## Outcome document
+
+060_outcome.md records, for a reader who was not in the loop: what landed and how,
+the run ids that prove it, what the tips caught that per-link CI would have caught
+earlier, what did not land and why, and the honest limits of the proof —
+specifically that non-tip pull requests merged without their own ci check under a
+recorded owner authorization.
+
+LOOP-PESSIMIST-01 applies to that document: it also records which hypothesis died
+and what evidence would show the tip-only CI economy is the wrong trade.
diff --git a/devlog/_plan/260913_contributor_carry_train/060_outcome.md b/devlog/_plan/260913_contributor_carry_train/060_outcome.md
new file mode 100644
index 0000000000..3cd0ca6548
--- /dev/null
+++ b/devlog/_plan/260913_contributor_carry_train/060_outcome.md
@@ -0,0 +1,123 @@
+# Outcome — the contributor carry train
+
+Twelve lanes were dispatched to land the open contributor work scored 60 or
+higher. Eleven landed, one needed nothing, and one is recorded separately below.
+
+## What landed
+
+| Wave | Lane | Tip | Merge commit | Carried |
+| --- | --- | --- | --- | --- |
+| 1 | L | #4481 | d865aacf93 | #4382 luvs01, #4413 rrmlima |
+| 1 | B | #4480 | 2af30c2d0e | #4381 luvs01, #4388 luvs01, #4460 AgenticLab-SH |
+| 1 | X | #4474 | 2296e485d6 | #4077 laerad777, copy residue only |
+| 1 | I1 | #4486 | a3ca64f605 | issues #4425, #4442 |
+| 1 | I2 | #4482 | 990cd8cce5 | issues #4430, #4435 |
+| 1 | C | #4487 | 55bb9f3fef | #4438 Yongzhaooo, #4389 olddonkey, #4457 jeongjin0 |
+| 1 | R | #4489 | 3f76ce415d | #4455 jeongjin0, #4409 yxr1995-maker, #4387 luvs01 |
+| 1 | S | #4477 | 981b53e7d0 | #4447 Veritas-7 |
+| 2 | I3 | #4500 | 94063d0798 | #4467 jaychou0642-create, issue #3775 |
+| 2 | I4 | #4498 | 8e6c99608c | issue #4454 |
+| 2 | H | — | — | nothing to carry; #3663 was already on dev |
+
+Every merge used the same gate: a Cross-platform CI run concluded success on the
+exact tip head SHA, the merge commit was verified with
+`git merge-base --is-ancestor` against `origin/dev` afterwards, and the
+`Co-authored-by` trailers were read out of the landed commits rather than the
+pull request bodies.
+
+## Dispositions
+
+Closed with evidence naming the landing and merge commit: source pull requests
+#4382, #4413, #4170, #4381, #4388, #4460, #4077, #4438, #4389, #4457, #4455,
+#4409, #4387, #4171, #4086, #4465, #4467, #3663, #4447, plus the carry links
+GitHub did not auto-close (#4479, #4473, #4485). Issues #4425, #4442, #4430,
+#4435, #4439, #4456, #4412, #3729, #4436, #3775 and #4454 were closed the same
+way.
+
+Nine further contributor pull requests were closed before the train started,
+because the previous 36-PR batch had already absorbed them: #4319, #4310, #3652,
+#4229, #4216, #4119, #4317, #4080 and #3458.
+
+Six issues stay open with a status comment naming what landed and what remains:
+#4191, #4311, #3661, #3781, #4312 and #3506. Each one has merged work that
+references it and no merged work that closes it, which is the distinction the
+previous batch learned to make.
+
+## What the reviews caught
+
+Every audit round in this train found something real, which is the argument for
+running them rather than trusting a green tip.
+
+The roadmap itself failed its first audit on five blockers, including two lanes
+prepared as peers that both write `src/server/responses/core.ts`, and a lane
+claiming a Windows CI leg that only runs on `workflow_dispatch`.
+
+The dispatch packets failed their first audit on five more. The severe one was
+an unqualified "never merge" in the common frame, which would have stopped every
+lane from running the `git merge origin/dev` the roadmap requires at a drifted
+tip. Another would have had lane X cherry-pick a commit that re-breaks an
+evidence-based model exclusion.
+
+Lane S took two security rounds. The first found that the canonical OpenAI seed
+defines only four keys, so "ignore keys the seed never defines" reached `headers`
+— which the PATCH mask writes and the forward adapter applies to the upstream
+ChatGPT request ahead of incoming headers. A dashboard-session PATCH would have
+ridden every later request.
+
+Lane I4 took three. Rounds one and two each found the same defect wearing a
+different payload: combo children bypassed the repair on their own cloned body,
+and the matcher required a well-formed Fernet token so near-miss ciphertext fell
+straight back into the original path. Round two then found the fix had traded
+fail-open for data loss — `looksLikeBackendCiphertext` is length ≥ 64 over a
+character class that a SHA-256 digest matches exactly. The landed version keeps
+the two slot kinds asymmetric: an `encrypted_content` slot is stripped whatever
+it holds, free text is matched strictly.
+
+## The lesson that repeated
+
+Three planned carries turned out to be already satisfied on dev: #4170 in lane L,
+#4086 in lane R, and the whole of lane H. All three were found by attempting the
+work, not by reading the plan.
+
+The inverse happened twice. #4465 proposed the #4442 fix before lane I1 existed
+and was not in the scored inventory, because the inventory is a snapshot taken
+before it was opened; the landing carries a trailer for its author. #4467 was
+caught the same way, but at dispatch time rather than after the fact, because the
+first incident turned into a standing check.
+
+Both directions are the same defect in the plan, not in the lanes: a snapshot of
+the open queue is stale the moment it is taken, and only the lane touching the
+code can tell.
+
+## Honest limits of the proof
+
+Non-tip pull requests merged without their own `ci` check, under the recorded
+owner authorization for tip-only CI. What makes that defensible is that each lane
+is cumulative, so the content of every link is a strict subset of what its tip's
+green run executed; the evidence exists, attached to the tip.
+
+No local full test suite was run at any point. Every suite claim traces to a
+hosted run id.
+
+Most `dev` runs triggered mid-batch ended `cancelled` as the next merge
+superseded them inside the concurrency group. That is the workflow behaving as
+configured, and the regression evidence is the completed runs on batch-containing
+commits rather than those cancelled ones.
+
+## LOOP-PESSIMIST-01: what did not improve
+
+The hypothesis that died is that a scored inventory plus a written roadmap is
+enough to dispatch from. It was wrong twice in each direction, and the correction
+was not a better inventory — it was giving every lane the obligation to
+re-verify its own inputs before writing code.
+
+What did not improve is lane-thread observability. `list_threads` is capped at
+50 and the newest lane tasks fell outside it repeatedly, so reaching a lane
+required reading session files off disk to recover its thread id. Coordination
+worked anyway, but it worked around the tool rather than through it.
+
+What would show the tip-only CI economy is the wrong trade: a defect landing on
+`dev` that a per-link run would have caught and the cumulative tip run did not.
+No instance appeared in this batch or the previous one. That is not proof it
+cannot happen — a lane whose links conflict semantically rather than textually
+could still produce one — and it is the specific thing to watch for.