Estimate token usage for vscode.lm providers that report none

[wch]

When Posit Assistant runs in Positron and a user selects a GitHub Copilot model, requests go through VS Code's vscode.lm API, which reports no token usage at all. As a result the Assistant showed no token counts for those conversations. This adds client-side token estimation for any vscode.lm provider that doesn't report usage: it counts tokens locally with the model's own countTokens() (a local tokenizer, no network call) and synthesizes the same end-of-stream usage event the rest of the pipeline already consumes. The synthesized usage is tagged as estimated via providerMetadata so consumers can display it as approximate.

Estimation triggers whenever a stream ends without having emitted a real usage part, so it covers Copilot today, extends to any other no-usage vscode.lm provider, and steps aside automatically if a provider ever starts reporting real usage. Positron's own providers, which already emit exact usage, are unaffected.

Changes

• src/positron/token-estimation.ts (new): input estimation from the final post-transform payload (messages + tool definitions + system parameter where applicable) with a per-message framing constant and a bounded module-level LRU memo keyed on model id plus a hash of the serialized message; output estimation from accumulated text and tool-call inputs. Exposed as a lazy thunk so providers reporting real usage never pay the countTokens cost.
• src/positron/VscodeLmClient.ts: wires the estimation context into the response-stream converter and, at stream end with no usage part, synthesizes a finish-step carrying the estimated usage and the providerMetadata.positai.usage.isEstimated = true marker. Estimation never fails the request — errors degrade to no-usage with a logged warning. The synthesized finish reason (and the existing real-usage one) now reflects tool-calls when the stream emitted a tool call.
• Tests: a fake LanguageModelChat with a scripted stream and deterministic countTokens covers estimation emitted for no-usage streams, suppressed when real usage arrives, the Copilot system prompt counted exactly once, memoization, error fallback, and the tool-call finish reason.

Not included

The core plumbing and UI that surface this flag (the ~ prefix, unpriced cost policy, TUI/web displays) live in the companion posit-dev/assistant PR, which pins to this branch.

Accuracy note

These counts come from the Copilot extension's local tiktoken tokenizer, so they are approximate for non-OpenAI models served through vscode.lm and are always presented as such. They deliberately do not apply any per-model correction factor. Anthropic's newer models (Opus 4.7/4.8, Sonnet 5, Fable 5) ship a tokenizer that produces roughly 1.0–1.35× as many tokens as the prior generation, but that ratio is old-Anthropic to new-Anthropic — not tiktoken to Anthropic — so applying it here would layer one guess on top of an unrelated one. Genuinely accurate counts for those models would require counting with the provider's own tokenizer (e.g. Anthropic's count_tokens), which is future work rather than a multiplier.