Stream the OpenAPI preview for Graph-sized spec selections - #1751

Previewing an OpenAPI integration with a spec-format selection (the Microsoft Graph preset URLs) parsed the 43MB Graph source whole. That builds a ~300MB tree, kills the 128MB Workers isolate mid-request, and surfaces in the add dialog as `Couldn't load or parse this spec` with an empty 503 behind it (trace `6af4828926a96ee93f9a608c7042e747`, 2026-08-26). The add/update path already streams via `structuralSplit`; the preview path did not.

• `previewSpecTextStreaming`: streaming twin of `previewSpecText` on the same structural-split primitive — head + schema-free components parse small, path-items parse one at a time through `keepPathItem`, and only the top-ranked health-check candidates get their response schemas walked, against the transitive `$ref` closure instead of the full schema map. `previewSpec` and the addSpec derived-preview branch on `resolved.keepPathItem`, so headless adds take the same path.
• Whole-document parses are now guarded by parsed-tree size: block YAML by line count (structure density is what OOMs — a 36MB flat-scalar doc still parses), JSON by text size, plus a coarse any-shape text ceiling enforced before download via Content-Length. Oversized generic specs get an actionable error instead of taking down the isolate.

Verified against the real 43MB Graph spec under a 160MB heap: mail preset previews 290 ops in ~0.8s at 129MB peak; the default 12-workload bundle previews 6,350 ops in ~1.0s at 154MB; the old whole-parse path now fails cleanly (`964,402 lines, limit 400,000`) instead of dying. Parity between streaming and whole-document previews is pinned by test on a shared fixture.