docs: document ModelOption.THINKING behaviour for Granite

[planetf1]

ModelOption.THINKING already works for controlling Granite 4.2 reasoning on the native Ollama, OpenAI, and LiteLLM backends — including the "low"/"medium"/"high" string levels — but nothing in the docs said so. The one substantive treatment that existed (openai.md's thinking section) was framed entirely around Qwen3 on vLLM using raw extra_body/enable_thinking, which pointed readers at the lower-level workaround instead of the higher-level ModelOption.THINKING API that's actually the preferred entry point when a backend honours it.

This expands the vague one-line ModelOption.THINKING table row in configure-model-options.md into a proper subsection with a per-backend behaviour table (native Ollama / OpenAI-compatible / HF), documents Granite 4.2's "low"-only effort distinction ("medium"/"high" are accepted but behave like True) and its default-on-when-unset behaviour, and adds a working example (docs/examples/thinking_mode.py) exercising True/False/"low" against a local granite4.2:3b Ollama model. It also cross-links openai.md's Qwen3/vLLM section to point at ModelOption.THINKING as the preferred portable mechanism, keeping the existing extra_body content as the documented fallback for backends/params it doesn't cover.

The LocalHFBackend row is documented as "not supported" (string values are silently dropped today) with a link to #1636, since the fix (#1639) hasn't merged yet. That row should be updated once #1639 lands.

What's out of scope

• The vLLM reasoning-parser callout and the configure-model-options.md/backends-and-configuration.md duplication noted in the issue — backends-and-configuration.md's table row already links elsewhere and wasn't called out in the issue's proposed changes, so it's left as-is.
• result.thinking parsing on LocalHFBackend (fix(hf): parse <think> tags into mot.thinking on LocalHFBackend #1616, open PR) and the HF string-drop fix (fix(hf): forward string THINKING levels as reasoning_effort, not low_effort #1639, open PR) are both separate in-flight PRs; this PR only documents current behaviour and links to them.

Verification

• uv run python docs/examples/thinking_mode.py — ran live against granite4.2:3b on local Ollama. Confirmed THINKING=True produces a long reasoning trace, THINKING="low" a short one, THINKING=False none, matching what's documented.
• npx markdownlint-cli "docs/docs/**/*.md" --config docs/docs/.markdownlint.json — clean on both edited files (one pre-existing, unrelated MD060 finding on a table this PR didn't touch).
• uv run ruff format/ruff check/mypy on the new example — clean.
• pre-commit run on the changed files — all hooks pass.

Fixes #1638

[planetf1]

Updated to document the `LocalHFBackend` row assuming #1639 (HF string-level forwarding) and #1616 (HF `` tag parsing) merge — both now describe HF as matching Ollama/OpenAI behaviour rather than "not supported".

Merge-order note: this PR should merge after #1639 and #1616, otherwise the `LocalHFBackend` row describes behaviour that isn't live on `main` yet.

[planetf1]

Code review — 3-reviewer panel, all findings addressed

Ran an independent tier-1/tier-2/tier-3 panel against this diff. All three reviewers verified and confirmed the same central finding independently, plus several narrower ones. Pushed a follow-up commit (f1a53834) fixing all confirmed items:

WARNING (all 3 reviewers, confirmed against source) — the OpenAIBackend/LiteLLM False cell claimed "Disables thinking" unconditionally. Verified against openai.py:747-749 (real OpenAI server type never gets reasoning_effort="none" — that branch is explicitly guarded out) and litellm.py:428-439 (only sent when "ollama" is in the model-id prefix). So False does not disable thinking for real OpenAI reasoning models or non-Ollama LiteLLM targets. Fixed: the cell now spells out the real-OpenAI/non-Ollama exception.

WARNING — the LocalHFBackend claims (string forwarding, result.thinking population) describe the state once #1639/#1616 merge, but the doc prose had zero inline signal of that — only the PR description said so. Fixed: added an inline > **Depends on #1639 and #1616** callout so the doc itself doesn't silently go stale if merged out of order.

WARNING — the unresolvable bare #1617 reference in openai.md's new callout (a published-site reader can't follow it). Fixed: reworded.

WARNING — no > **Full example:** link to docs/examples/thinking_mode.py, per CONTRIBUTING_DOCS.md's convention, leaving the inline snippet to drift from the real file. Fixed: added.

SUGGESTIONs fixed: "same mechanism as the OpenAI backend" was inaccurate (HF forwards through a gated chat-template variable; OpenAI sends an ungated top-level param) — reworded; "This holds across all three backends" now notes the runtime-forwarding dependency; added a non-determinism note to the new code block per doc convention; added cheap assertions to the example pinning the documented per-arm behaviour (reran live against granite4.2:3b, all pass).

NITs fixed: dropped the unnecessary qualitative pytest marker (no quality assertion existed at the time it was added); fixed US-English spelling (behaviour → behavior) in new content per CONTRIBUTING_DOCS.md's house style.

Not addressed (correctly scoped FOLLOW-UP, pre-existing, unrelated to this diff): missing **See also:** footer and a missing .md extension on an unrelated pre-existing link in configure-model-options.md.