chore: bump stale model references to current-generation values (examples, auto.go, docs)

[aheritier]

What

A proactive sweep to bump stale/outdated model references to current-generation
values across examples/*.yaml, pkg/config/auto.go + auto_test.go, and
provider/docs pages. Follow-up to #4121
(models.dev snapshot refresh, unrelated/already merged) — this is a separate,
dedicated arc.

Policy

Keep two generations live per model family where the provider actually
has two majors: the latest major's newest minor, and the previous major's
newest minor. Never an arbitrary older minor. When a family has two valid
anchors and a given file uses neither, split by cost/fallback signal in the
key/file name: keys/files signalling cost or fallback (budget,
budget-model, fast, discounted-sonnet, claude_fast) get the previous
anchor; everything else gets the latest anchor. Already-valid anchor values
are left untouched.

Established anchors (live-verified against real provider catalogs)

Family	Latest (current major)	Previous (prior major)
Anthropic Sonnet	claude-sonnet-5	claude-sonnet-4-6
Anthropic Opus	claude-opus-5	claude-opus-4-8
Anthropic Haiku	claude-haiku-4-5 (single, no split)	—
Anthropic Fable	claude-fable-5-1 (single, no split)	—
Bedrock Sonnet/Opus	global.anthropic.claude-{sonnet-5,opus-5}	global.anthropic.claude-{sonnet-4-6,opus-4-8}
OpenAI flagship	gpt-5.6-sol	gpt-4.1
OpenAI mini	gpt-5.6-terra	gpt-4.1-mini
OpenAI nano	gpt-5.6-luna	gpt-4.1-nano
Google flash	gemini-3.8-flash	gemini-2.5-flash
Google flash-lite	gemini-3.5-flash-lite	gemini-2.5-flash-lite
OpenRouter Llama	meta-llama/llama-4-maverick	meta-llama/llama-3.3-70b-instruct
Moonshot Kimi	kimi-k3	n/a
xAI Grok	grok-4.5 (single default)	—
Baseten DeepSeek	deepseek-ai/DeepSeek-V4-Pro	deepseek-ai/DeepSeek-V3.1

Judgement call: docs/providers/anthropic/index.md's claude-opus-4-7 row

That row's wording ("Previous Opus flagship; supports task budget") is
literally documenting the still-supported previous generation, and
claude-opus-5 already occupies the "highest-capability" row directly above
it — so the row is bumped to claude-opus-4-8 (the previous-generation Opus
anchor), not claude-opus-5. This differs from examples/task_budget.yaml,
which has no such generation-labelling context and no cost/fallback signal on
its opus-bounded/opus-bounded-tight keys, so it moves to claude-opus-5
(the latest anchor) instead.

Scope

• examples/*.yaml — ~60 files, mapping below.
• pkg/config/auto.go DefaultModels + the mirroring assertions in
  pkg/config/auto_test.go.
• docs/providers/*/index.md and docs/configuration/flavors/index.md.

examples/*.yaml (old → new)

See the implementation PR (linked below) for the full per-file diff; summary
of the anchors applied: every claude-sonnet-4-5 → claude-sonnet-5 (except
files signalling a cheaper/fallback tier, which get claude-sonnet-4-6),
every plain claude-opus-4-6 → claude-opus-5, claude-fable-5 →
claude-fable-5-1, every bare gpt-4o/gpt-5 used as a "just give me a
capable model" default → gpt-5.6-sol, gpt-4o-mini/gpt-5-mini →
gpt-5.6-terra, grok-4.3 → grok-4.5, kimi-k2-0905-preview → kimi-k3,
meta-llama/llama-3.3-70b-instruct → meta-llama/llama-4-maverick, dated
Bedrock/Anthropic pins dropped in favour of the bare anchor.

examples/task_budget.yaml and examples/thinking_budget.yaml needed
individual review (task budget semantics / thinking-effort tiers are
version-sensitive) — see the judgement call above and the PR description.

pkg/config/auto.go DefaultModels

Provider	Old	New
anthropic	claude-sonnet-4-6	claude-sonnet-5
google	gemini-3.5-flash	gemini-3.8-flash
openrouter	meta-llama/llama-3.3-70b-instruct	meta-llama/llama-4-maverick
baseten	deepseek-ai/DeepSeek-V3.1	deepseek-ai/DeepSeek-V4-Pro
moonshot	kimi-k2-0905-preview	kimi-k3
amazon-bedrock	global.anthropic.claude-sonnet-4-5-20250929-v1:0	global.anthropic.claude-sonnet-5

pkg/config/auto_test.go's mirroring assertions are updated identically
(mechanical, no independent judgement).

Explicitly excluded

• examples/fireworks.yaml / docs/providers/fireworks/index.md — the
  configured Fireworks model ID doesn't exist in any generation on
  models.dev; tracked separately in Fireworks default/example model reference (kimi-k2-instruct) doesn't exist in models.dev, silently skipped by CI #4132.
• examples/nebius.yaml — already removed via Remove the nebius example #4130 (merged).

Acceptance

• Every mapped reference bumped as described.
• task build, task test, task lint pass.
• Dedicated PR opened referencing this issue.