feat: show per-message token usage and estimated cost (USD) in chat

[Meinianda-L]

Closes #1703

Type of change

• New feature

What does this PR do?

Shows per-message token usage and estimated cost (USD) under each assistant
message in the chat, similar to the cost display in Cline/OpenCode. Issue
#1703 requested per-message token usage; its comments also asked for an
estimated cost.

Claude Sonnet · 16,000 tokens · $0.0960

Changes:

1. core: PromptLog now carries usage (and the raw model id) from the
   LLM request. llm/streamChat already passes the PromptLog through, so no
   protocol change is needed.
2. core: calculateRequestCost refactored into a shared cost helper;
   pricing lives in core/llm/utils/modelPricing.json, extracted from the
   models.dev API (https://models.opencode.ai/api.json — the same data
   source OpenCode uses at runtime), retrieved 2026-08-08, with the source
   attributed in the file. Prices extended to DeepSeek, Gemini and Mistral,
   plus cache-read pricing for GPT-4o/gpt-4o-mini, and gpt-3.5-turbo
   aligned with models.dev (0.5/1.5). Display-only, not
   billing-authoritative.
3. gui: new UsageCost footer component. The line is omitted entirely
   when no usage was reported; the cost part is omitted when the model price
   is unknown — we never show a made-up figure.

Not changed: the console/stats page and CLI (they already show cost), the
llm-info package, no runtime dependency on models.dev.

How did you verify your code works?

# core (node 20.20.1, jest)
NODE_OPTIONS=--experimental-vm-modules npx jest core/llm/utils/calculateRequestCost.test.ts
  17 pass, 0 fail

# core full suite: 856 pass (19 pre-existing failures in llm.test.ts require API keys)

# gui (vitest)
npm test
  377 pass, 0 fail

# repo
npx tsc --noEmit   # core + gui clean
npx eslint ...     # clean on changed files
npx prettier --check ...  # clean

New tests cover known-model pricing (Anthropic/OpenAI/DeepSeek/Gemini/Mistral),
unknown model/provider -> null, zero tokens, prefix/case matching, cache
write/read branches, and the GUI footer's render/omit behavior.

Screenshots / recordings

N/A — formatting is plain text (Model · 16,000 tokens · $0.0960).

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR

[github-actions]

All contributors have signed the CLA ✍️ ✅
_{Posted by the CLA Assistant Lite bot.}

[Meinianda-L]

Friendly ping for review: @sestinj and @bdougie。 The PR is self-contained and extends the existing cost calculation in core/llm/utils/calculateRequestCost.ts, so it can be reviewed and merged independently.

[Meinianda-L]

I have read the CLA Document and I hereby sign the CLA

[Meinianda-L]

Heads-up on the failing checks — neither is related to this PR:

1. test (windows-latest, 22): infra flake, npm registry returned 502 Bad Gateway - GET https://registry.npmjs.org/npm/-/npm-10.9.3.tgz during install. Could a maintainer re-run this job?

2. jetbrains-tests: pre-existing failure — Autocomplete > testAutocomplete() fails on main too (see merged commit badf887, same DriverWithContextError). This PR doesn't touch autocomplete or the IntelliJ extension (the only core change adds two optional fields to the chat PromptLog return value).

All other checks pass: lint, prettier, core-checks, gui-checks, vscode-checks, binary-checks, packages, e2e, and tests on 18/20/22/24 across all three OSes.