feat: price AI Gateway long-context requests at their tier prices - #30916

Bills an AI Gateway request at its model's long-context tier prices when the request's prompt is over the tier's threshold, instead of always at the base prices.

Stack context

AI Gateway prices every cache write at the 5-minute price and ignores the long-context prices that models.dev publishes, so Claude usage and long prompts are under-priced. On dogfood, Oct 1-14 recorded $7,535 against a list price of about $9,290. The 1-hour cache rule closes about 86% of that gap, and the tiers close almost all of the rest. The stack, bottom to top:

1. chore: add 1h cache write and long-context tier prices to the price book #30915: adds the 1-hour cache write and long-context tier prices to the price book.
2. feat: price AI Gateway long-context requests at their tier prices #30916 (this PR): bills long prompts at their long-context tier prices.
3. feat: price AI Gateway 1h cache writes #30872: bills 1-hour cache writes at the 1-hour price.

Why

Some models publish higher prices above a prompt size, for example Claude Haiku 5.5 above 100K tokens and GPT-5.6 and later above 272K tokens, on their own APIs and on Bedrock. AI Gateway billed those requests at the base prices. #30915 added the tier prices to the price book; this PR bills with them.

Pricing rule

• The prompt size of one upstream request is its input, cache read, and cache write tokens.
• When the prompt size is over a tier's threshold_tokens, every token of that request uses the tier's prices. A prompt equal to the threshold uses the base prices.
• With several tiers, the highest threshold below the prompt size applies. A price that the tier does not set uses the base price.
• Past usage is not repriced. On startup the seed adds tiers to default price rows; custom prices keep their own values.

What changed

• aibridge sends the prompt size on each token usage record (aibridged API v1.4), and structured token_usage logs include it as prompt_tokens. In a stream, the output-only message_delta record carries the prompt size of its message_start, because the two records reach coderd in no fixed order.
• Database (migration 000628): ai_model_prices.tiers (a jsonb array), and on aibridge_token_usages the prompt size and the threshold of the applied tier, so each row's price snapshot still adds up to cost_micros.
• Cost: pricesForPrompt in coderd/aibridgedserver/cost.go picks the tier, and the usage row stores the prices that applied.
• API, SDK, and CLI carry tiers. Validation rejects negative prices and thresholds that are not positive and strictly ascending. The CLI lists tiers as an optional column and shows tier changes in the update plan.
• Model form: one read-only line per tier under the cost estimate (story CostEstimateWithTiers).
• Docs: cost-controls.md states the rule.

What to scrutinize

• The prompt-size floor in usageFromRequest: max(prompt_tokens, input + cache read + cache write). Records from AI Gateway daemons on API v1.3 send no prompt size, so they still select a tier by their own input.
• A stored tiers value that does not decode fails the usage record instead of billing at the base prices. The API and the seed validate tiers, so only a manual database edit can cause this.
• UpsertAIModelPrices now also rewrites a default row when only its tiers change, so existing deployments pick up tiers on upgrade.

Generated files

All from make gen: coderd/aibridged/proto/aibridged.pb.go (protoc); coderd/database/queries.sql.go, querier.go, models.go, dump.sql, and check_constraint.go (sqlc and the schema dump); coderd/apidoc/* and docs/reference/api/schemas.md (swag); site/src/api/typesGenerated.ts.

Validation

• go test passes for aibridge/intercept/messages, aibridge/recorder, aibridge/internal/integrationtest, coderd/aibridgedserver, coderd/aibridge/prices/..., coderd/database/migrations, db2sdk, dbgen, and coderd/aibridged/..., and with -run AIModelPrice in enterprise/coderd and enterprise/cli.
• Site: tsc, Vitest (src/modules/aiModels and the ModelForm stories), Biome, oxlint, and knip pass.
• Each new test was checked to fail when the behavior it covers is removed.