docs(usage): clarify turn cost accounting

[any-victor]

Summary

• document reported per-response token counts, including absent and basic-usage limitations
• add a fresh exhaustive formatted-source audit of the exact pinned Cursor IDE 3.18.9 artifact
• distinguish RunInference tokens, DashboardService event, conversation/time-bounded and aggregate costs, quoted usage-based prices, Admin API usage events, and Agent SDK cost
• state why none of the available monetary contracts provides an exact identifier-level join to a Pi assistant message
• replace imprecise invent/agree/disagree wording with explicit support and equality language

Corrected source evidence

• pinned RunInference schema and formatted adapter: typed response arms contain tokens and invocationId, but no typed monetary field; provider metadata is opaque and the client does not read a billed-cost key from it
• pinned DashboardService schema: GetFilteredUsageEvents returns event rows with TokenUsage.totalCents, cursorTokenFee, chargedCents, model, timestamp, client type, and optional conversationId
• pinned DashboardService schema: GetClientUsageData accepts conversationId plus int32 timestampBeforeRequest and returns named costInCents items; full packaged-source searches find no caller or field consumer, so its unit, item meanings, settlement timing, and concurrency behavior remain unestablished
• pinned AiService schema and formatted workbench: CheckUsageBasedPrice is used only by the usage-based-pricing preflight display; it has no call identifier and is not post-response measured usage
• DashboardService monetary shapes have no RunInference invocationId, request ID, response ID, or generation ID
• pi-cursor uses one conversationId for the Pi session and a distinct invocationId for each inference call
• current /cursor usage calls only aggregate DashboardService methods; filtered/client endpoint account access remains source-verified rather than live-verified
• official Cursor Admin API exposes a closely matching hourly event shape; Agent SDK cost belongs to Cursor-managed agent runs

Verification

• exact artifact, agent-host chunks, and desktop workbench byte counts and SHA-256 hashes match the committed lock
• cost-field census and full packaged 3.18.9 caller/field searches completed; 3.18.25 retains the same contracts and caller boundaries
• mise run verify
• mise run changeset:status

No paid inference call was made.