fix(providers): add per-model attachment capability override (#2741)

[Sayt-0]

Problem

Custom and aliased OpenAI-compatible providers (xai, mistral, nebius, ollama, and user-defined proxies) resolve attachment capabilities through a models.dev lookup keyed on the config provider/model string (base.Config.ID()). When that pair is absent from models.dev, modelinfo.LoadCaps swallows the "not found" error and returns conservative text-only capabilities, so attachment.Decide drops every image and PDF part with only a generic warning. Affected models lose attachments silently.

Fixes #2741.

Verified against the live models.dev catalogue:

Provider string	In models.dev?	Failing case
ollama	no (only ollama-cloud, no local tags)	ollama/llava dropped
xai	yes, but grok-2-vision-1212 is gone (drifted to grok-4.x)	legacy vision model dropped
custom proxy	no (user-defined)	vision-capable gpt-4o dropped

Why not a provider mapping table

The issue listed a provider to models.dev mapping table as one option. Live api.json shows it would fix almost nothing: xai, mistral, nebius, minimax, github-copilot already match models.dev exactly (identity mapping), ollama has no catalogue entry at all, and neither model-absence (grok-2-vision-1212) nor user-defined providers can be fixed by a provider rename. The vertexai precedent (withModelsDevProvider) works there only because those models do exist in models.dev under a different key, which does not hold here.

What this does

Implements two of the issue's three options: the per-model escape hatch and the non-silent diagnostic.

• New typed override on the model config: capabilities: {image, pdf}. When set it is authoritative and skips the models.dev lookup; when omitted, behaviour is unchanged.
• A deduplicated diagnostic (once per model id) when a model is absent from models.dev, naming the id and pointing at the override.

Layer	Change
pkg/config/latest	CapabilitiesConfig plus optional Capabilities field on ModelConfig; isShorthandOnly updated
pkg/modelinfo	CapsOverride plus ResolveCaps(store, id, override); once-per-model miss diagnostic in LoadCaps
pkg/model/provider/base	Config.CapsOverride() builds the override from config
providers	override threaded through the 5 conversion paths: oaistream (openai chat, dmr), openai responses, anthropic, bedrock, gemini
agent-schema.json, examples/capability-overrides.yaml	schema plus example

modelinfo.ResolveCaps takes plain booleans, so modelinfo keeps no dependency on the config package (layering preserved). The override is resolved once per client and passed down; a nil override reproduces prior behaviour at every call site.

Example:

models:
  llava-local:
    provider: ollama
    model: llava
    capabilities:
      image: true
      pdf: false

Issue expectations mapping

Issue option	Status
Per-model capability override in config	done (capabilities)
Warn rather than silently drop	done (deduplicated miss diagnostic)
Provider to models.dev mapping table	not done; shown ineffective against current models.dev data

Backward compatibility

Capabilities is optional and nil by default; existing configs are unaffected. Only pkg/config/latest is touched (frozen config versions untouched).

Tests

Test	Covers
pkg/model/provider/oaistream repro	the drop for an uncatalogued provider/model, and the override restoring the image
pkg/modelinfo ResolveCaps plus diagnostic	override precedence, nil fallback, once-per-model warning
pkg/config/latest	YAML round-trip, shorthand, deep clone of the override
pkg/model/provider end-to-end	override survives Registry.New to the built client CapsOverride()
pkg/model/provider/base	CapsOverride() mapping

Notes and residual design points

• Status is needs-design; the chosen direction is the override plus the diagnostic. The field name and the "override fully wins" semantics are open for maintainer review.
• Anthropic keeps the "any Claude model supports image and PDF" heuristic when no override is set; for a Claude model absent from models.dev the diagnostic may fire once even though the heuristic then restores capabilities. This is harmless and deduplicated.