docs: add thinking/reasoning guide and expand provider thinking docs

[rumpl]

New docs/guides/thinking/index.md covering thinking/reasoning across all providers, plus expanded sections on OpenAI and AWS Bedrock.

• New guide (docs/guides/thinking/): cross-provider reference — effort levels, token budgets, adaptive thinking, interleaved thinking, thinking display, task budget, and a complexity-based chooser table
• Bedrock: add Thinking Budget and Interleaved Thinking sections with effort-to-token mapping table and beta header behaviour (previously only a one-liner in a table)
• OpenAI: replace one-liner with effort-level table, hidden reasoning token callout, and adaptive thinking example
• nav.yml: register the new guide under Guides

See examples/thinking_budget.yaml for a runnable demo.

[docker-agent]

Assessment: 🟡 NEEDS ATTENTION

2 documentation inconsistency findings in added lines. Both relate to the OpenAI max effort level being omitted from standalone tables while adaptive/max is referenced in YAML comments — creating a potential user-facing ambiguity.

[docker-agent]

[MEDIUM] max effort level missing from standalone table but referenced in adaptive comment

The effort-level table (lines 64–71) lists none, minimal, low, medium, high, xhigh — but the Adaptive Thinking YAML comment on this line includes adaptive/max, implying max is a recognised level. This creates an internal inconsistency: if max is a valid OpenAI effort level, it should appear in the standalone table (and in the inline comment on line 59). If it is only valid inside adaptive/, this asymmetry may silently mislead users who try thinking_budget: max without the adaptive/ prefix.

Suggested fix: Either add max as a row in the effort table (between xhigh and a note that it is rarely needed), or explicitly note in the adaptive section that adaptive/max is a special-case form not available as a bare level.

[docker-agent]

[MEDIUM] Quick Reference table omits max for OpenAI while adaptive comment references adaptive/max

The Quick Reference table on this line lists OpenAI's valid values as minimal, low, medium, high, xhigh, none, adaptive/<level> — with no mention of a bare max level. Yet the adaptive YAML comment on line 72 explicitly enumerates adaptive/max as a valid choice. Compare with the Anthropic row (line 28) and AWS Bedrock row (line 31), both of which list max explicitly.

If max is a valid standalone OpenAI level (as it was in the old inline comment and as it maps to a token budget for Bedrock), it should appear in this table. If it is intentionally excluded from the non-adaptive path, a brief note would prevent users from copying the Anthropic pattern and wondering why thinking_budget: max is rejected.

Suggested fix: Either add max to the OpenAI values cell, or add a footnote clarifying that max is only valid as adaptive/max for OpenAI.

[rumpl]

Addressed both reviewer comments (comments #3360724109 and #3360724113):

• Confirmed via pkg/effort/effort.go (ForOpenAI): only minimal, low, medium, high, xhigh are valid standalone levels — max is intentionally excluded.
• Added a note in both the thinking guide and the OpenAI provider page clarifying that max is only valid as adaptive/max, not as a bare thinking_budget value.