feat(pricing): add generic selector pricing and Responses cache writes

[Menci]

What

Add standard Responses cache-write accounting and a generic pricing-selector mechanism used for context-length and service-tier prices.

• model usage.input_tokens_details.cache_write_tokens, split disjoint ordinary/cache-read/cache-write usage with one shared invariant checker, and throw on malformed overlapping totals
• map Responses cache writes to Messages cache_creation_input_tokens in streaming and non-streaming translations
• define pricing as flat explicit cells with a generic selector map
• centralize selector axes, price-vector validation, canonicalization, catalog compilation, authoring helpers, runtime projection, and exact lookup in @floway-dev/protocols/common
• ship serviceTier open-string equality and inputTokens gt/gte threshold axes
• require threshold bands to be declared without equality coordinates before service-specific combinations can reference them
• project the highest matching global threshold, then exact-match the canonical Cartesian coordinate
• leave missing coordinates unpriced by design; there is no cross-cell fallback, precedence, inheritance, or multiplication
• compile stable pricing catalogs once with a WeakMap; static provider tables eagerly validate through shared pricingCell / modelPricing / basePricing helpers

Canonical selector identity is sorted-key JSON; threshold objects serialize as { "operator", "value" }, and the base coordinate is {}.

Persistence and transfer

Migration 0052_usage_pricing_selector.sql replaces per-axis usage columns with one JSON-object-validated pricing_selector column. Existing nonblank tiers migrate through SQLite json_object; NULL/blank tiers become {}. SQL and memory repos canonicalize typed selector identity.

Data-transfer format v10 exports the self-describing pricingSelector; imports validate selectors and non-negative rates and reject historical versions.

Current threshold prices

All thresholds reprice the whole request.

• GPT-5.6 Sol / Terra / Luna: standard short and inputTokens > 272000 long; Codex also has priority-short. Priority-long is intentionally absent/unpriced.
• Copilot GPT-5.5: >272000 = 10 / 1 / 45 input/cache-read/output.
• Copilot and Codex GPT-5.4: >272000 = 5 / .5 / 22.5.
• Copilot Gemini 3.1 Pro Preview: >200000 = 4 / .4 / 18; Gemini 2.5 Pro remains base-only because the current Copilot catalog caps it below the threshold.
• Ollama MiniMax M3: base .3 / .06 / 1.2; >512000 = .6 / .12 / 2.4.

Authoring UI

The model editor iterates exported axis metadata. Equality axes render free text; threshold axes render operator + integer controls. Catalog-level validation errors render once above the cell list, while invalid in-progress cells remain safe to edit. Threshold clearing, row changes, and manual/auto mode switches are covered.

Custom upstream catalog pricing remains best-effort: malformed cost metadata is omitted without dropping the model.

Test plan

• pnpm run typecheck
• pnpm run lint
• pnpm run test — 337 files / 4084 tests pass
• canonical identity, selector/rate validation, duplicate/conflicting coordinates
• global threshold-band invariant, gt/gte boundaries, highest-match projection
• exact-missing combinations and unknown runtime tiers remain unpriced
• shared provider authoring helpers eagerly validate static catalogs
• SQL/memory selector identity and 0052 blank/special-character migration
• v10 export/import and historical-version rejection
• generic Vue controls, threshold clear/invalid drafts, one-time error rendering, row/mode resync
• shared Responses cache-write splitting and malformed-usage exceptions
• malformed custom catalog cost omission

[Menci]

Claude Code agent update: this PR is paused for a broader architecture comparison. The strict >272k whole-request math and explicit priority×context cells are verified, but I am now comparing how LiteLLM, models.dev, Portkey, Helicone, OpenRouter, Cloudflare and other gateways represent context-length pricing alongside service tiers. I will not treat the current inputLengthTiers/input_above_tokens design as final until that multi-implementation review is complete.

[Menci]

Claude Code agent architecture result after comparing LiteLLM, models.dev, Portkey, Helicone, Langfuse, OpenRouter, and Cloudflare's public contract:

• context length is a pricing selector independent of wire service_tier; it is not another token billing dimension and not a subtype of service tier
• crossed thresholds select a whole-request price, not marginal overage
• selector combinations are often sparse explicit cells; missing cells must not be auto-multiplied or silently inherit an unrelated axis
• persistence should retain tier and input_above_tokens as separate bucket coordinates

The current nested inputLengthTiers[].tiers shape still privileges one axis. I am refactoring this PR to a symmetric flat list of explicit selector coordinates and complete price vectors. Empty selector is the base cell; exact missing combinations remain unpriced. The existing migration coordinates and cache-write accounting stay.

The review also found that the new coordinate changes backup identity, so the export/import format will be bumped to version 9 rather than silently extending version 8.