docs(devlog): record the lane I catalog and model-visibility unit

devlog/_plan/260918_lane_i_catalog_visibility/010_lane_record.md

@@ -0,0 +1,82 @@
+# Lane I — catalog and model visibility
+
+Six assigned issues, worked against `dev` at `a0f611d4aceb9476d44268e43722273b7b211846`.
+Hosted CI was the only executable verification; no local suite, typecheck, build, install or
+`ocx` invocation was run in this lane.
+
+## Disposition
+
+| Issue | Outcome | PR |
+| --- | --- | --- |
+| #4811 | Sync-log suppression diagnostic plus the two missing config keys | #4958 |
+| #4940 | Proxy-side refusal of a Reserve turn the opt-in cannot serve | #4968 |
+| #4944 | Generated-registry data refresh for `opencode-go` | #4962 |
+| #4646 | Doctor check for an unexposed pinned default model; two asks answered, not implemented | #4963 |
+| #4590 | Base prompt reported by the prompt probe | #4964 |
+| #4662 | Client-role discriminator so a missing management route is legible | #4959 |
+
+Filed from this lane: #4961 (the `codexDesktopAuthless` coupling defect), #4971 (the latent
+`seen`-skip), #4972 (the shared default loopback port).
+
+## Findings worth keeping
+
+**#4811 and #4940 are not one defect.** They fail at two different clauses of the same guard.
+`createReserveCatalogProjection` refuses on
+`!isEffectiveCodexDesktopAuthless(config) || mainSelectors.length === 0`. #4811 has the opt-in on
+and trips the second clause; #4940 has it off and trips the first, never reaching the selector map
+at all. The working hypothesis going in was that both were the selector map, and two patches would
+have been written against it.
+
+**#4944's operative cause was not the mechanism its title leads with.** The title names three, and
+the `seen` skip in `routed-gather.ts` is the one that reads like the culprit. It is not: the
+serialized entry gets its window from `applyCatalogMetadata`
+(`src/codex/catalog/parsing.ts:828-836`) straight from the generated table, with no reference to
+the `CatalogModel` list, so a stale registry is the cause and a merge fix would have changed no
+row. Recorded as #4971 so the ruling-out survives the closure. The issue's model list is also
+partially wrong: `muse-spark-1.3-contributor` was already declared at 1048576 in the provider
+registry preset, and `deepseek-flash` is not published by models.dev for this provider.
+
+**#4646 ask 1 was already correct behaviour.** Sync byte-compares the regenerated catalog against
+disk rather than testing for file existence; the guard landed 2026-08-11 in
+`c7eec01ca4`/`642805c11e`, a month before the report. Implementing the requested drift trigger
+would have rewritten the catalog on every sync forever, because `visibleNativeSlugs` omits
+disabled slugs from `/v1/models` while the catalog deliberately retains them as `hide` rows — two
+sets that are unequal by design — and that rewrite would re-break the #857/#1407 mtime staleness
+contract.
+
+**#4662's reported cause was a red herring.** There is no second management port; management rides
+the same listener. The ambiguous 404 came from the client machine listener, which publishes a
+`role` discriminator that the CLI liveness parser was discarding.
+
+## Pattern: the file-size ratchet makes large files append-only
+
+Two lanes hit this independently, so it is worth naming rather than rediscovering.
+
+`tests/fixtures/file-size-baseline.json` caps named files, and `scripts/file-size-ratchet.ts`
+treats any growth past the cap as `GREW`. The baseline only moves downward — `updateBaseline`
+uses `Math.min` — so a file sitting exactly at its cap cannot be extended at all, and raising the
+cap is not an available move. That is deliberate: a lane that reaches for the cap has already
+decided to mask something.
+
+The consequences in practice:
+
+- `tests/codex-integration/codex-catalog.test.ts` sits exactly at 7985. The #4944 regression could
+  not be added to it. The sanctioned escape is a new sibling file plus its two registrations, in
+  `scripts/test-layout/layout.json` `explicit` and `tests/fixtures/test-layout-expected.json` —
+  the shape `d3ca5522db` established, and what `catalog-hub-context-window.test.ts` already is.
+- `src/codex/catalog/sync.ts` is capped at 52 and was at 52. A six-line facade re-export added in
+  #4958 pushed it to 58 and failed a Linux shard. The fix was to delete the re-export, which was
+  never load-bearing, rather than to look for a way to move the cap.
+
+Check the baseline before planning where a change lands, not after CI says no. A one-line node
+read of `tests/fixtures/file-size-baseline.json` against the current line counts catches it in
+seconds, and it is worth doing across every file a lane intends to touch.
+
+## Correction made during implementation
+
+The #4646 brief asserted that a disabled-but-hidden native slug "will fail at request time". It
+does not. `src/router.ts:922` states that routing never consults `disabledModels`, and
+`docs-site` `model-routing.md:80` already documented that disabling does not reject a direct
+request. The slug is routed by the ordinary rules as though it were enabled. The shipped docs and
+the issue's closing comment state the accurate behaviour; the brief was wrong and the
+implementation caught it.