fix(ci): repair the weekly model-limits drift check and refresh model data

[justin-carper]

Summary

The scheduled model limits · drift vs Cursor docs job has failed on every weekly run (latest: 2026-09-28). This fixes the job and brings the model data it guards back in line with Cursor's docs.

Root cause

scripts/sync-model-limits.mjs fetchDoc sent accept: text/plain,text/markdown,*/*. cursor.com returns 404 for that header and 200 with default fetch headers, so --check exited 2 (HTTP 404) before reading any data. Dropping the header is the one-line fix; the rest of this PR is what the check then surfaced.

Changes

• Generator
  • The pricing doc now has two tables with identical columns (Cursor Models pool, then Other Models). The parser reads every matching table. A model with a row in both is still reported as ambiguous, not first-wins.
  • MODEL_IDS now matches the live 40-id catalog. Added: claude-fable-5-1, claude-opus-5-5, claude-sonnet-5-5, gemini-3.7-flash, gemini-3.8-flash, grok-4.6, grok-4.7, muse-spark-1.3. Dropped: glm-5.2 (absent from the live catalog).
  • The grok-4.5 and composer-2.5 $0 "pool" overrides are removed: the docs now list price rows for both. auto-smart, default and composer-2 keep $0 as placeholder rates, with corrected why strings.
  • The NO_AUTO_COMPACTION_INPUT_LIMIT block was hand-inserted into the generated src/model-limits.ts in 27d11bb without updating the template, so a regen would have deleted it. It now lives in the template.
• src/model-limits.ts: regenerated with npm run sync:model-limits, not hand-edited.
• Cost changes for existing models (from the docs): claude-sonnet-5 $3/$15 → $2/$10, gpt-5.6-sol $5/$30 → $4/$20, grok-4.5 $0 → $2/$6, composer-2.5 $0 → $0.5/$2.5. No context window changed.
• vitest.config.ts (new): excludes .cursor/**. The plugin mirrors skills into <cwd>/.cursor/skills/, and default discovery then collects their test files (94 load failures when running npm test in a dogfooded checkout).
• Tests: added a two-table generate test that fails if the parser goes back to first-table-only; removed two tests that only echoed generated cost numbers and assert the invariant instead.

Verification

• npm run sync:model-limits -- --check: exit 0, prints up to date (context: 36 from docs, 4 overridden | cost: 36 from docs, 4 overridden | 40 model ids), which is the line ci.yml greps for
• npm run typecheck, npm run build: pass
• npm test with no --exclude flags: 42 files / 642 tests pass
• Six generated values spot-checked against both live docs pages by a reviewer
• Drift job dispatched on this branch via workflow_dispatch (run 36711158492, re-run on the rebased head 6565d6d): model limits · drift vs Cursor docs success (it had failed on every scheduled run before this)

Independent of #132 (different files apart from CHANGELOG.md [Unreleased]; whichever merges second needs a trivial rebase).