docs(#2872): provider × capability matrix — single source of truth

[Skobeltsyn]

ChatGPT audit (#2865) P1: capability claims are currently scattered across CHANGELOG entries, multimodal.md, model-and-tools.md, and per-adapter KDoc. A single canonical matrix makes drift visible.

New docs/providers.md — the four columns (Anthropic / OpenAI / Ollama / DeepSeek) match ModelProvider.entries. Six matrices:

• Modality input — image is shipped on Anthropic / OpenAI / Ollama; DeepSeek shape-tested only; document / audio / video deferred (see the #2868 erratum for the v0.6.5 release-note slip).
• Reasoning — Anthropic / Ollama / DeepSeek surface text; OpenAI is count-only.
• Prompt caching — wire-level vs automatic-prefix vs engine-KV; which fields land on TokenUsage.
• Tool calling — ToolChoice.{Auto, Required, None, Specific} wire shapes per provider; Ollama's best-effort hint + JUL warning.
• Constrained decoding — all four ship supportsConstrainedDecoding() == true with provider-specific wire shapes.
• Streaming — SSE / NDJSON, what chunk types each provider emits.

Closing "Updating this matrix" section names the 4-step checklist for adding a fifth provider (enum + adapter + matrix column + caching tests + caching.md update).

Cross-linked from:

• README.md "Documentation" list (between Skills and Model & Tool Calling).
• docs/multimodal.md Stage 1 paragraph (modality scope).
• docs/model-and-tools.md lead-in (capability scope vs DSL scope).

Auditor flagged the scattered claims as: "Agents.KT sells trust. Trust products need unusually clean docs." This concentrates the trust surface into one file.

No code changes. Full ./gradlew test detekt green.