devlog: work-phase records for the backlog disposition program

devlog/_plan/260822_backlog_disposition_program/090_wp9_new_pr_disposition.md

@@ -0,0 +1,122 @@
+# 090 — WP9: disposition of the four PRs that arrived mid-loop
+
+#2361, #2362, #2363, #2364 all arrived from one contributor while work-phase 0 was
+running, and all four claim to close an issue. Four independent read-only review lanes
+were dispatched (`xai/grok-4.6`, high reasoning effort, after the
+`openrouter/stealth-ox-alpha` lanes hit provider 429s and were retired).
+
+## Verdicts
+
+| PR | Claims | Verdict | Disposition |
+|----|--------|---------|-------------|
+| #2361 | Closes #2356 | GO-WITH-FIXES (blockers=0) | **MERGE** |
+| #2362 | Closes #1809 | reviewer lane failed to return; **reviewed directly** | **LEAVE OPEN**, blockers restated |
+| #2363 | Closes #1225 | **FAIL** | **LEAVE OPEN**, blockers restated |
+| #2364 | Closes #1406 | **FAIL** | **LEAVE OPEN**, blockers restated |
+
+## #2361 — merge
+
+The only one that does what its issue asked. The reviewer confirmed the maintainer's
+three requirements: the sentinel is converted inside `mapReasoningEffort`, openai-chat's
+existing `undefined` branch does the omitting, and **no default omission was baked into
+the ollama registry entry** — which the triage explicitly forbade.
+
+Load-bearing proof, run in the reviewer's own throwaway worktree: reverting
+`src/reasoning-effort.ts` to the merge base makes the new test fail with
+`expected undefined, received "__omit__"`. The test is real.
+
+Three non-blocking findings accepted as-is: the sentinel is undocumented in
+`src/types/provider.ts` and docs-site (Medium, discoverability), a dead post-clamp
+sentinel check that can never fire, and an unused exported helper
+`isReasoningEffortOmitted`. None changes behavior.
+
+This also supersedes **#2357**, the wrong-base draft against `main` for the same issue.
+#2357 closes rather than retargets — the branch policy forbids feature PRs against
+`main`, and a maintainer rewriting a contributor's PR base is worse than asking them to
+reopen.
+
+## #2363 — FAIL, and the proof is the interesting part
+
+The reviewer deleted the **real call site** in `writeRetainedCatalogSync` and re-ran the
+PR's own tests:
+
+```
+(pass) applyAutoReviewModelOverride sets auto_review_model_override across all entries
+(pass) applyAutoReviewModelOverride is a no-op when autoReviewModel is null or empty
+  2 pass, 0 fail
+```
+
+The tests pass with the feature disconnected. They exercise the exported mutator, never
+the write path — so CI would stay green while catalog sync emits `null` again.
+
+Two more blockers stand independently:
+
+- **The dashboard writer never stamps.** `prepareCatalog` / `convergeCodexCatalog`
+  (`src/codex/convergence.ts:367`) rebuilds routed rows from templates with
+  `auto_review_model_override: null`. The GUI path undoes what `ocx sync` just wrote,
+  which is the original #1225 failure mode returning through the primary surface.
+- **No slug validation.** Issue #1225 requires validating the target against the same
+  sync's catalog and failing clearly on an unresolved target. A stale slug is stamped
+  silently, and fail-closed auto-review then denies every approval.
+
+Also: native rows are overwritten without the opt-in the issue asked for, and docs are
+absent.
+
+**#2041 is superseded regardless.** The reviewer found it calls an undefined
+`configuredAutoReviewModel()` — a runtime `ReferenceError` — and bumps `package.json`
+on top of being CONFLICTING. #2363 is the better vehicle even though it is not yet
+sufficient. #2041 closes in wp8.
+
+## #2364 — FAIL
+
+Commit 1 wired management validation and `safeConfigDTO`; **commit 2 deleted both.** The
+reviewer proved the consequence live against the PR head:
+
+```
+mgmt invalid   null          <- schema-invalid body accepted
+dto.vercelGatewayRouting  undefined   <- valid config hidden from GET /api/config
+```
+
+So `POST /api/providers` can persist config that `loadConfig` later rejects, salvaging
+away the whole provider. OpenRouter — the direct parallel this PR models itself on —
+validates at exactly that site. Docs are also absent, which #1406 explicitly required.
+
+One reviewer finding is worth recording as *dismissed*: CodeRabbit asked for the payload
+under `providerOptions.gateway`. Vercel's Chat Completions documentation accepts the
+top-level `provider` shorthand, which is what the issue and the maintainer asked for.
+The AI reviewer was wrong; the PR is right on that point.
+
+## #2362 — reviewed directly after a failed dispatch
+
+Its lane produced nothing across three wait cycles and was retired under
+DISPATCH-RETIRE-01 rather than waited on indefinitely. Reviewed by the main agent
+instead.
+
+It adds three new config keys — `modelResponsesCompatibility`,
+`modelResponsesTerminalRepair`, `responsesTerminalRepair` — to
+`src/types/provider.ts` and reads them in `src/providers/registry.ts`. The changed-file
+list is:
+
+```
+src/providers/registry.ts
+src/types/provider.ts
+tests/deepseek-inbound-wire.test.ts
+```
+
+Neither `src/config.ts` nor `src/server/auth-cors.ts` appears. Comparable per-model keys
+on `dev` are validated in both — `modelAdapters` has
+`modelAdapterRecordConfigError` at `src/config.ts:1463` **and**
+`src/server/auth-cors.ts:615`. This is the same structural gap #2364 was failed for,
+so it gets the same disposition rather than a pass by luck of which lane returned.
+
+The escape-hatch logic itself is gated on `effectiveAdapter === "openai-responses"` and
+falls through to the registry policy, so it is opt-in and does not change default
+behavior — the design is sound. It is the config-surface wiring that is missing.
+
+## Why three PRs are left open rather than closed
+
+Each carries real, correctly-diagnosed intent from a contributor who found genuine
+problems. Closing them would discard that over fixable gaps. Each gets its blockers
+restated on the PR with the exact evidence, so the author can finish the work — which is
+the outcome the repository actually wants.
+