feat(models): add input-modality overrides for catalog-gap vision models

[eibednejo]

Related to #90 — same root cause as #91, the other capability. #91 covers the missing reasoning flag; this covers missing input modalities.

Problem

src/commandcode-catalog.ts is generated from a single Command Code CLI release. A model published afterwards is absent from MODEL_INPUT_MODALITIES, so inputModalitiesForModel() falls back to ["text"] — and both transports honor that absence:

• index.ts:127 publishes input: [...inputModalitiesForModel(model.id)] to pi and Oh My Pi, so the host never offers image attachments for the model.
• src/core.ts:543 calls assertTextOnlyMessages() on the generate transport, so an image in a user message fails the request outright with Selected Command Code model does not support image content in user messages.

commandcode-catalog-overrides.ts only exposed MODEL_EFFORT_OVERRIDES, so there was no supported way to restore image input for a catalog-gap model.

Evidence

command-code@1.53.0 (dist/cli.mjs) declares image input for both models:

id:"deepseek/deepseek-v4.1-flash",inputModalities:["text","image"],…,description:"V4.1 hybrid-attention reasoning with vision",…
id:"xai/grok-4.6",inputModalities:["text","image"],…,reasoning:!0,…

The generated catalog (from command-code@1.44.0) has no entry for deepseek/deepseek-v4.1-flash at all, and xai/grok-4.6 appears only in MODEL_REASONING and MODEL_EFFORTS — neither is in MODEL_INPUT_MODALITIES. Both ids are served by GET /provider/v1/models.

Change

• src/commandcode-catalog-overrides.ts — add MODEL_INPUT_MODALITIES_OVERRIDES with its own doc block (when to add, when to drop) and the two entries.
• src/models.ts — merge the overrides over CATALOG_MODEL_INPUT_MODALITIES, mirroring how MODEL_EFFORTS already merges MODEL_EFFORT_OVERRIDES.
• tests/test-models.ts — test mirroring the existing effort-override test: each override must not be upstream yet (drop signal), must include image, must reach inputModalitiesForModel()/modelSupportsImageInput(), and generated entries must still pass through unchanged.
• CHANGELOG.md — Unreleased entry.

Verification

• npm run test:unit — 217/217 pass; typecheck and prettier --check clean.
• Live on a Go subscription, with the generated catalog left untouched and no user-level models.json override in play: deepseek/deepseek-v4.1-flash correctly described a test image (red circle, blue square, green triangle, text ZEBRA-42). Before this change the same request fails with the text-only assertion quoted above.

Notes

• xai/grok-4.6 could not be exercised end to end: it returns MODEL_NOT_IN_PLAN on a Go plan (GOAT and above required). Its entry rests on the CLI registry declaration — the same source npm run sync:commandcode-catalog reads.
• Two further ids declare image upstream but are absent from GET /provider/v1/models, so they are deliberately not added: gpt-6-astra and minimax/minimax-m3-free.
• This touches the same files as feat(models): enable reasoning for deepseek-v4.1-flash #91 (src/models.ts, src/commandcode-catalog-overrides.ts, tests/test-models.ts, CHANGELOG.md). Happy to rebase on whichever lands first.
• npm test exits 1 in my environment on tests/test-pi-local.mjs ("Claude request through Anthropic Messages endpoint", 404). It reproduces identically on unmodified main, so it is unrelated to this change; CI is authoritative for the suites that need the full toolchain.

[patlux]

Thanks for identifying the missing image metadata. Merged #98 now includes upstream image modalities for DeepSeek V4.1 Flash and Grok 4.6, and #95 repairs the sync automation. This supersedes the per-model override entries without adding another maintenance path. The independent host-input behavior in #94 remains planned. Replacement CI passed before merge.