Every engine fetch goes through the one HTTP loader: retry, abort and named errors

[pasquelin]

Why

Split from #457 (several pull requests). The engine has one HTTP loader, cluster/pages.ts:12 checked (one retry on a network error or 5xx, an EngineError naming the URL), yet five places fetch on their own with other policies (architect, 2026-09-25): no retry, plain Errors, no abort signal.

To do

• One loader, one policy: checked gains optional (a 404 answers null) and credentials; gpu/page/pages.ts:183 httpPageSource, host/prepared/images.ts:29, lighting/importedLights.ts:36, physics/tiles.ts:52 and physics/tilePlace.ts:49 call it; the physics tiles take the model's abort signal.
• The changed error codes documented: docs/SDK.md, the API reference and every translation.

Code context

cluster/pages.ts:12 and the five call sites above.

Proof

Tests per call site: 404, 503 then 200, abort. Fast gates green.

Links

Split from #457.

[pasquelin]

taken by lead sdk

[pasquelin]

CTO decisions (2026-09-26), relayed by lead sdk: (1) for optional reads only, a 403 counts as absent like a 404 (S3 and CloudFront answer 403 for a missing object); a required read fails on either. (2) streaming/fetch.ts joins checked in this pull request if it stays near 500 lines, else it is the first To-do on #457; either way it stops retrying a 4xx.

[pasquelin]

Path note (lead sdk): the To-do's credentials option is not added: its only use would have been 'same-origin' in host/prepared/images.ts, which is already fetch's default, so it would have no user. streaming/fetch.ts already went through checked; it now stops at a 4xx (decision B).