Add CI workflow to validate the exercise dataset

[Mavi9412]

Summary

This repo currently has no automation at all — every data or translation PR is merged on trust. This adds a single GitHub Actions workflow that validates data/exercises.json on every PR and push to main, covering five checks:

• JSON Schema validation — data/exercises.json against data/exercises.schema.json (draft 2020-12, via ajv)
• No duplicate ids
• Every image path resolves to a real file
• Every gif_url path resolves to a real file
• index.html's embedded copy of the dataset stays in sync with data/exercises.json — index.html inlines the full dataset as a JS literal for offline use; today it's byte-identical to data/exercises.json, but nothing currently stops that from silently drifting on a future data PR

The workflow installs ajv/ajv-formats at CI time (npm install --no-save) rather than committing a package.json/lockfile, to keep the repo's zero-build, "just open the HTML files" design intact.

Why

The last several merged PRs in this repo's history are contributors adding new-language instructions (French, Hindi, Polish, Korean...). That pattern is exactly where an unenforced schema causes real friction — daceea8 is a follow-up commit fixing the JSON Schema's language maps after a language PR had already been merged once. This workflow would have caught that at review time instead of needing a second commit.

Design notes

• The index.html sync check parses the embedded array with a small bracket-depth scanner rather than a fixed regex, so it keeps working whether the array is minified or reformatted to multi-line JSON later.
• Parse/compile failures (a malformed exercises.json, a broken schema) are caught and reported as a clear ✗ message, not a raw stack trace — since a JSON syntax slip is the most likely mistake for a contributor hand-editing a ~150k-line file, and it deserves the clearest error of anything this workflow checks.
• push is restricted to branches: [main] and pull_request covers incoming PRs (including forks), so a PR from a branch in this repo doesn't trigger the workflow twice for the same commit.
• permissions: contents: read and timeout-minutes: 5 are set explicitly — the job only reads the repo and doesn't need more than that.

Test plan

Verified locally by deliberately breaking each check one at a time against the real dataset and confirming a correct failure, then restoring:

• Valid dataset passes all five checks
• A duplicate id is caught
• A missing image file is caught
• A missing gif_url file is caught
• A schema violation (bad body_part enum, missing required language) is caught
• index.html going out of sync with data/exercises.json is caught
• Malformed JSON in data/exercises.json produces a clean error, not a stack trace
• A broken schema (invalid keyword) produces a clean error, not a stack trace
• A record missing image/gif_url entirely reports a clear message instead of "missing image file \"undefined\""
• The workflow YAML parses correctly and the push/pull_request path list stays in sync via a YAML anchor