[Bug]: Catalog advertises image input for text-only models when visionSidecar.enabled=false (2.84.0) #7086

Problem

On 2.84.0, the exported Codex catalog advertises text,image for a provider-configured text-only model even when the global vision sidecar is explicitly disabled. The corresponding request-time planner returns no sidecar plan. This makes the client attachment affordance inconsistent with the effective configuration.

This is a static-source report plus a read of an existing locally exported catalog, not a newly executed model request or isolated runtime reproduction. The installation has other local patches, but all five source files cited below are byte-for-byte identical to the pristine 2.84.0 npm package.

Minimal relevant configuration

{
  "visionSidecar": { "enabled": false },
  "providers": {
    "deepseek": {
      "noVisionModels": ["deepseek-v4-pro"],
      "modelInputModalities": { "deepseek-v4-pro": ["text"] }
    }
  }
}

The provider block is an excerpt, not a standalone complete configuration. No credentials or endpoints are needed to describe the mismatch. The existing exported entry for deepseek/deepseek-v4-pro has input_modalities: ["text", "image"].

Source evidence in published 2.84.0

• src/codex/catalog/model-hints.ts:296: sidecarCovered = isModelVisionSidecarConsumer(prov, model.id) widens the model to image input.
• src/codex/catalog/routed-gather.ts:747: the custom-row path uses the same predicate and widens to image input.
• src/vision/eligibility.ts:136: this predicate only receives provider capability declarations and a model id. It cannot see the global disabled setting.
• src/vision/plan.ts:222: the actual planner returns undefined when visionSidecar.enabled === false. The function documentation specifies stripping raw images when no plan is available.
• src/codex/catalog/gather-capture.ts:567: gatherFlightKey does not include the global sidecar-enabled setting, so a fix should also consider sharing/caching across configuration snapshots.

Expected behavior / scope

When the vision sidecar is explicitly disabled, do not add synthetic image capability to otherwise text-only rows. Preserve native image capability and explicit supported overrides, and preserve sidecar-backed image advertisement when enabled. I am not asking to enable a sidecar automatically or send images to a text-only upstream.

Related history: #349 and #3105 deliberately enabled attachment delivery for sidecar-backed models. This report concerns the explicitly disabled-sidecar case. If advertising attachments while stripping them with no plan is intentional, please clarify the intended contract and whether a supported setting can advertise effective text-only capability without disabling a model.

No proxy/client restart, live model request, authentication changes, or TLS changes were made to investigate this report.