feat(ai): move chat models to Opus/Sonnet/Haiku 5.5

[ABB65]

What changed

The Studio chat catalog moves to the 5.5 generation. This covers Studio chat and the ee Conversation API: request params, pricing and credits, compatibility for stored IDs, and handling of refusals.

• shared/utils/ai-models.ts: the catalog, the default model (DEFAULT_CHAT_MODEL = 'claude-sonnet-5-5', :199), the new ModelPricing.longPrompt card (:151) and resolveChatModelId (:220).
• shared/utils/ai-credits.ts: legacy pricing kept for old IDs (:77-79), ratesForPrompt (:129), and estimateMessageCostUsd (:158), which now prices the long-prompt part on its own card.
• server/utils/turn-budget.ts: ratesForEstimate (:108); TurnUsageTracker.pricedUsage(model) (:256) classifies each call against the threshold, including the call in flight.
• server/utils/conversation-engine.ts: the refusal branch (:497). The done event carries longPrompt.
• server/providers/ai.ts:160,183: 'refusal' stop reason.
• chat.post.ts:163 settles from tracker.pricedUsage (still in finally). chat.post.ts:246 resolves the requested ID before the plan gate.
• ee/enterprise/conversation-api.ts:422-483: carries the per-call long split into the settle.
• server/utils/conversation-keys.ts:73,84 and ee/enterprise/conversation-keys.ts:48,109,187: key list, create and update resolve stored and sent IDs.
• app/composables/useChat.ts:253-262: a retired ID in the picker's localStorage moves to the current one and is saved again. :759 shows the refusal message.
• .contentrain/content/system/ui-strings/en.json: the new chat.refused key, edited in the JSON directly with no MCP write.
• supabase/migrations/051_conversation_key_model_default.sql: new migration (see below). 001_baseline.sql is untouched.

Old → new model IDs

Stored / sent	Resolves to
claude-haiku-4-5-20251001, claude-haiku-4-5, any claude-haiku-*	claude-haiku-5-5
claude-sonnet-5, claude-sonnet-4-*, claude-3-5-sonnet-*, any claude-sonnet-*	claude-sonnet-5-5
claude-opus-5, claude-opus-4-*, any claude-opus-*	claude-opus-5-5
anything without a catalog family (gpt-4o, claude-fable-5-1, ...)	null, so the caller's existing default fallback applies

Pricing (USD per MTok, from the live pricing page)

Model	Input	5m write	1h write (Studio)	Cache read	Output
Opus 5.5	4	5	8	0.20	20
Sonnet 5.5	2	2.50	4	0.10 (0.05x)	10
Haiku 5.5, prompt ≤ 100K	0.10	0.125	0.20	0.01	0.50
Haiku 5.5, prompt > 100K	0.50	0.625	1	0.05	2.50

• The 1h write is 2x input on all three models, so CACHE_WRITE_MULTIPLIER = 2 stays.
• Sonnet 5.5 cache read: the live page (platform.claude.com/docs/en/about-claude/pricing) says $0.10 (0.05x). The founder facts and the migration guide said $0.20. This PR encodes the page's $0.10. A question for ORK is below.
• Haiku tier: the prompt length counts all input tokens (fresh + cache read + cache write), each request is priced on its own, and the higher rates apply to output too. The code models this per call: the turn tracker keeps a per-call list, pricedUsage splits out the calls over 100K, and the settle, turn budget and Conversation API all use that split. When a summed total arrives without a split, it is classified as a single request. A multi-call turn over 100K in total is then priced on the long card, which is above real cost and never below.
• Old IDs keep their pricing (claude-sonnet-5 $2/$10, claude-haiku-4-5* $1/$5), so past usage rows still price correctly.

API-breaking changes handled

• Thinking: every catalog model gets thinking: {type: "adaptive"} + an explicit effort (server/providers/anthropic-ai.ts:254-270, requestThinking). disabled is sent only for IDs outside the catalog. This fixes Sonnet 5.5's 400 on disabled, and Haiku 5.5 never gets budget_tokens.
• Effort: Haiku low, Sonnet medium, Opus medium, always sent explicitly so the recalibrated defaults (Sonnet high, Haiku medium) never apply.
• tool_choice / sampling: no request path sends tool_choice, temperature, top_p, top_k, prefill or fallbacks. A unit test asserts this for every catalog model (tests/unit/anthropic-ai.test.ts). The engine is the only request path for both Studio chat and the Conversation API.
• Preserved thinking: history replays byte-identically (existing invariant). Thinking blocks travel through block_binding: drop_block as before. Persisted rows are not rewritten.
• Refusal: Haiku 5.5 has no server-side refusal fallback. stop_reason: "refusal" now ends the turn with an error event, code: 'refused' (conversation-engine.ts:497), and runs no tools. The UI shows chat.refused. Credits still settle in finally from real usage.
• Sonnet 5.5 max output: raised from 16K to 32K, because adaptive thinking counts toward it. Haiku stays at 16K, Opus is unchanged.
• Prompt cache layout: unchanged. No per-request content went into system.

Stored-ID compatibility

• Studio chat: chat.post.ts:246 resolves body.model before the plan gate. A retired ID becomes the family's current model. An unknown ID falls back to the default as before.
• Chat picker: useChat reads the stored ID through resolveChatModelId and saves the new ID straight away.
• Conversation keys: the list, create and Conversation API runtime use resolveConversationModel, which resolves and then requires a non-premium API model (haiku-5-5 / sonnet-5-5), else falls back to the default. Update writes ai_model only when the sent ID resolves.
• Migration 051 changes only the column default of conversation_api_keys.ai_model to claude-sonnet-5-5. It rewrites no rows, because read-time resolution covers them. It is one lineage: the next number after 050, applied by both runners.

Credit and gate impact (plan prices and limits unchanged)

• Haiku 5.5 is about 10x cheaper than Haiku 4.5. Most Haiku turns now settle at the 1-credit floor. Starter economics improve.
• Sonnet 5.5 has the same input/output price as Sonnet 5 and a cheaper cache read. It now thinks adaptively at medium (Sonnet 5 ran thinking-off), so output tokens per turn will rise.
• Opus 5.5 is unchanged. The premium gate still closes Opus in a trial on the Studio key.
• Credit weights are unchanged. The AI_CREDIT_UNIT_USD floor holds: settle prices at or above real cost in every path.

Questions for ORK

1. Sonnet 5.5 cache read: this PR encodes the live page's $0.10. If a contract or invoice says $0.20, change one number (cacheReadMultiplier: 0.05 → 0.1 in ai-models.ts).
2. Credit floor on Haiku: the 1-credit-per-message floor now sits well above the real cost of a light Haiku 5.5 turn. That is fine for "never below unit", but you may want to revisit it.
3. Starter/free limits and the trial premium gate: worth a look now that Haiku is about 10x cheaper. Nothing changed here.
4. Effort tuning: Haiku low and Sonnet medium were not checked against evals (no live calls in this PR).

Known limits

• A call cancelled mid-thinking undercounts output, because thinking is not streamed as text. This gap existed before, and now applies to all three models.
• The history budget estimator undercounts Haiku 5.5's tokenizer (about 30% more tokens). The Haiku history budget stays at 24K.

Tests

Suite	Files	Tests
unit	200 passed, 1 skipped	2287 passed, 1 skipped
integration	54 passed	517 passed
nuxt	52 passed	358 passed

• pnpm lint: 0 errors. The 7 v-html/console warnings were already there.
• pnpm typecheck: clean.
• Integration needs the generated #contentrain client (contentrain-query generate). Without it, the billing webhook suites fail on import.
• No live API calls.