[CloudCorpRecords]

Fixes #199.

Root cause

npx convex codegen runs through startComponentsPushAndCodegen (src/cli/lib/components.ts), which builds a startPushRequest and calls startPush -> POST /api/deploy2/start_push to get the backend's evaluated schema for the generated bindings. The request carried dryRun: options.dryRun, and for a normal npx convex codegen invocation that is false, so the backend provisioned the schema as part of start_push and created the tables and indexes on the deployment. The CLI only ever needed the evaluated analysis; it never finalizes a push in the codegen path (no finishPush). The --help text saying codegen "doesn't modify the code running on the deployment" was therefore wrong in practice.

The change (src/cli/lib/components.ts)

• The start_push request is now marked as a dry run whenever the push is a codegen run (dryRun: options.dryRun || options.codegen, extracted into the exported isDryRunStartPush() helper). dryRun is the protocol's established don't-mutate signal, the same one npx convex deploy --dry-run relies on; the evaluated schema the bindings need is exactly what a dry-run push returns.
• The "Uploading functions to Convex..." spinner is codegen-aware now ("Analyzing functions for codegen...") since codegen no longer uploads anything.
• Added src/cli/lib/components.codegen.test.ts with a truth-table test pinning the policy: normal deploy is not a dry run, deploy --dry-run is, and codegen always is.

Test evidence

• npx vitest run: 6/6 passing (5 existing aiFiles tests + the new codegen test).
• tsc --noEmit --emitDeclarationOnly false: clean, exit 0.

Honest limits

The table/index creation happens server-side on start_push; I could not run codegen against a live deployment from this sandbox, so the backend's dry-run behavior is taken from the protocol's documented contract (deploy --dry-run previews without applying). Happy to adjust if the team prefers a dedicated flag over reusing dryRun.