Read model prices and context windows from LiteLLM instead of hand-kept tables

[calvintvu]

Context

we dont want to manually sync model information, we'll follow ccusage pattern and pull info from litellm

Changes

• The collector reads prices and context windows from one embedded table, generated from LiteLLM (MIT). A new model release needs no code change.
• The collector computes each cost and sends it with the session. The frontend holds no pricing table.
• make models rewrites the table. The Models workflow runs it daily and opens a pull request, so a person reads every price change before it merges.
• The workflow compares the model list before and after each refresh. The pull request names every model added and dropped, and a dropped model turns the run red.
• Three prices now match LiteLLM: gpt-5.6-terra input is $2.00 per MTok, not $2.50. gpt-5.6-luna input is $0.20, not $1.00. The gpt-5.5-pro cache read is $3.00, not $30.00.
• The context window of claude-sonnet-5 is 1,000,000, not 967,000. The context gauge shows a smaller percentage for those sessions.
• LiteLLM carries no entry for claude-mythos-5 or gpt-5.5-cyber. Sessions on those models show the unpriced-model warning and no cost.
• A Codex session takes its context window from the table when the transcript reports none.

Test

• make check - passed. The collector carries no test files, so make test runs nothing.
• npm run format:check, npx tsc -b, npm run lint, npm test (8 tests), npm run build - all passed
• Manual: ran the dev binary and read /api/sessions. 191 sessions, total cost $1166.56, no unpriced model, subagent costs present, claude-sonnet-5 window 1,000,000.
• Manual: ran the workflow comparison locally. A real refresh reported no change. A table with claude-opus-5 deleted reported the drop and exited 1.
• The last commit deletes o4-mini-2025-04-16 from the table on purpose, to watch the workflow report a model as added. Revert it before merge.

Screenshots

[calvintvu]

6e190b8 removes model o4-mini-2025-04-16 from models json, upon merge we can test workflow to see if we updated the models json correctly

[luhe19001]

This backend migration regresses cost and context-window handling for model IDs that were supported by the previous pricing table but are absent from the generated LiteLLM subset. For sessions using previously supported IDs such as claude-mythos-5, gpt-5.5-pro, or gpt-5.5-cyber, the lookup now fails after only removing a date suffix, so usage becomes unpriced and contributes $0. claude-mythos-5 also loses its context window. Please preserve local aliases or fallback entries for these IDs when moving pricing to the backend.