feat(llm): make output limits and provider beta features tier-configurable

[devlikebear]

Part of #919. Phase 0 of the provider modernization roadmap. LP-004.

Prerequisite: #922 (LP-003) — the thinking budget must fit under whatever max_tokens becomes configurable here.

Summary

Add max_tokens and provider beta feature flags to the tier binding, and plumb them through the resolver and router so frontier-model capabilities are reachable from config.

Problem

Output is pinned at 4096 with no way to raise it. The Anthropic client defaults to 4096 when unset:

// internal/llm/provider.go:373
if config.MaxTokens <= 0 {
    config.MaxTokens = 4096
}

and nothing ever sets it. LLMTierBinding (internal/config/types.go:141) and ResolvedLLMTier (internal/config/llm_resolve.go:22) have no max_tokens field, and helpers_llm_router.go:34 builds ProviderOptions without one — so MaxTokens is structurally always 0. Every tier is capped at 4096 output tokens regardless of model.

This also collides with LP-003: a thinking budget must be strictly below max_tokens, so any meaningful budget is unreachable while the cap is 4096.

Beta features are unreachable. anthropic-beta is one hardcoded constant:

// internal/llm/anthropic.go:16
const anthropicPromptCachingBeta = "prompt-caching-2024-07-31"

sent unconditionally at anthropic.go:67. Prompt caching went GA long ago, so this occupies the only slot with a header that no longer does anything — while 1M context, interleaved thinking, and extended cache TTL all need that slot.

It is also sent to third-party Anthropic-compatible gateways (config/default.yaml:21 points minimax at one), which may reject or mishandle an unknown beta.

Scope

• Add max_tokens to LLMTierBinding and ResolvedLLMTier; pass it through helpers_llm_router.go.
• Replace the hardcoded beta constant with a list assembled per request.
• Add a tier-level opt-in for provider beta features, named provider-agnostically so other providers can adopt the mechanism.
• Drop the GA'd prompt-caching beta from the default set.
• Choose per-model output defaults instead of a blanket 4096 when a tier omits max_tokens; keep 4096 as the last-resort fallback for unknown models.
• Validate thinking_budget < max_tokens at resolve time with a loud error, not at request time.
• Document the new fields in config/tars.config.example.yaml.
• Surface the fields in the console tier editor and the onboarding wizard review step.

Acceptance Criteria

• A tier can produce a response longer than 4096 tokens.
• A tier with no max_tokens still works, using a sane per-model default.
• thinking_budget >= max_tokens fails at config load with a message naming both values.
• Beta flags are opt-in per tier and absent from the request when unset.
• A kind: anthropic gateway tier can be configured to send no beta header at all.

Validation

• Resolver unit tests for present/absent/invalid max_tokens
• Unit test for the thinking_budget vs max_tokens guard
• Request-builder tests asserting the assembled beta header for: default, opt-in, and explicitly empty
• Round-trip test through YAML → env override → resolved tier
• make test
• make lint-diff

Out of Scope

• Context-window budgeting — that is LP-006, which lands on the same tier binding and should be sequenced after this.
• Per-role output limits. Tier granularity is enough for now.