fix(capabilities): resolve the MiMo v2.6 family as vision-capable (#14587)

[xiechimon]

What

Refs #14587. The MiMo v2.6 family (mimo-v2.6-pro, mimo-v2.6-flash) is natively vision-capable but absent from every capability source in this repo: the opencode-go registry stops at mimo-v2.5, the MiMo block in modelSpecs.ts covers only the v2.5 generation, and VISION_MODEL_ID_FRAGMENTS has mimo-vl but nothing for v2.6. resolveVisionCapability therefore returns null, and both /v1/models and the combo projection render the model text-only. Because combo capability is an intersection over members, a single unresolvable member flips the whole combo to multimodal: false.

(Changed from Fixes to Refs per review: the issue still tracks the two out-of-scope gaps below, so it should not auto-close from the release branch.)

Why this fix

Adds "mimo-v2.6-pro" and "mimo-v2.6-flash" to VISION_MODEL_ID_FRAGMENTS (src/shared/constants/visionModels.ts), the shared last-resort heuristic from #4072 that routing, the /v1/models catalog and lite compression all delegate to. This is the minimal fix the report proposes, and the established channel for exactly this case (same pattern as vision-models-cc-fragments.test.ts covers for gpt-5, kimi-k2., claude-fable).

Per review, the fragments are scoped to the documented pro/flash family instead of the whole mimo-v2.6 generation, so a hypothetical mimo-v2.6-distill-qwen-9b or mimo-v2.6-tts does not inherit the vision verdict; the comment beside the fragments says how to extend it if Xiaomi documents image input for more v2.6 ids. The fragments match both provider-qualified ids from the report (opencode-go/mimo-v2.6-pro, command-code/xiaomi/mimo-v2.6-pro) plus bare and :free variants. They cannot over-match the text-only earlier-generation pro siblings: mimo-v2.5-pro and mimo-v2-pro never contain v2.6, and their hard text-only overrides (isKnownTextOnlyDespiteSync, checked first in the cascade) stay untouched.

Also rescopes the stale "MiMo *-pro = text-only" comment in modelSpecs.ts to the v2.5 family: the v2.6 generation flips that rule (models.dev lists mimo-v2.6-pro with image input). Comment-only change, no spec entries touched.

Explicitly out of scope, per the report's own framing as separate fix gaps: (a) a first-class per-model vision override, (b) per-member resolved supportsVision in /v1/combos. Both are design calls beyond this fragment fix.

How this was tested

New regression test tests/unit/14587-mimo-v2-6-vision.test.ts (16 cases):

• Positive: the two ids from the report, bare mimo-v2.6-pro, :free suffix, mimo-v2.6-flash, -ultraspeed all resolve supportsVision === true through both isVisionModelId and getResolvedModelCapabilities.
• Negative: mimo-v2.5-pro, mimo-v2-pro (bare + provider-qualified) stay non-vision, mimo-v2-flash untouched, unrelated text model control, plus the two undocumented-sibling cases from the review (mimo-v2.6-distill-qwen-9b, opencode-go/mimo-v2.6-tts) stay non-vision.
• Resolved-capability controls: mimo-v2.5 and mimo-v2-omni still true, earlier-gen pro verdicts unchanged.

Red-green (re-run after the narrowing): with both src files reverted to base release/v3.8.51 the 6 v2.6 vision cases fail (supportsVision resolves falsy), the 10 others pass; after the fix 16/16 pass.

Adjacent regression: vision-models-cc-fragments, vision-detection-consistency, mimo-v2-5-safety, command-code-registry-vision = 56/56 after the change; the earlier wider sweep (combo/capability suites 94/94, all vision/mimo/models tests 1057 pass / 0 fail / 3 skipped, the 8 combo-suite failures identical on a clean base worktree = pre-existing) predates the narrowing and the touched code path is the same shared fragment list. prettier --check clean, eslint exit 0 on the three touched files.

No changelog fragment, following the precedent of #14300 / #14291.

[diegosouzapw]

Thanks @xiechimon, clean and well-tested fix. Your new test fails on the release tip and passes here, and Xiaomi's model page confirms V2.6-Pro takes image input.

Two small things:

1. Please change Fixes #14587 to Refs #14587 in the description. The release branch is our default branch, so Fixes would auto-close the issue, and it still has the two parts you marked out of scope (the ignored compat supportsVision override and the per-member combo diagnostic).
2. Optional: mimo-v2.6 also matches ids like mimo-v2.6-distill-qwen-9b or a future mimo-v2.6-tts. If you'd like, narrow it to the pro/flash family (or exclude distill/tts) and add one negative case.

[xiechimon]

Both done in f7ac29d.

1. The description now says Refs #14587, with a line noting why.
2. Took the narrowing: the fragment list now carries mimo-v2.6-pro and mimo-v2.6-flash instead of the whole generation, so mimo-v2.6-distill-qwen-9b and mimo-v2.6-tts stay non-vision. Added both as negative cases in the regression test (16 cases now), and re-ran the red-green: with the src files reverted to base the 6 v2.6 vision cases fail, after the fix 16/16 pass. The comment beside the fragments explains how to extend it if Xiaomi documents image input for more v2.6 ids.