fix: SSR error prod-silence keys on the dev flag, not NODE_ENV

[vivek7405]

Summary

Closes #483.

defaultSSRErrorTemplate (the per-component SSR error isolation) and the core renderToStream boundary catch gated prod-silence on NODE_ENV. But webjs keys prod on the CLI dev flag (the deployment docs say so), and webjs start does NOT export NODE_ENV=production, so a bare prod launch leaked a thrown component's error message (a DB error, a stack-derived path) to the client. This also made #482's "the two streaming error paths now agree" only true when NODE_ENV=production was set.

The fix threads the server dev flag through the SSR render context: ssr.js stamps it on the suspenseCtx; renderToString / renderToStream source it (opts.dev wins, else the ctx) and back-fill the ctx so a streamed sub-render inherits it, then thread it to injectDSD and the boundary catch. Prod-silence now triggers whenever dev is false, regardless of NODE_ENV. A context-free renderToString with no dev signal falls back to NODE_ENV, preserving the prior behaviour for a bare unit/embedded render. The component-isolation path, the page-streaming boundary, and the core renderToStream boundary now agree on the same signal.

Test plan

• Integration (packages/server/test/dev/ssr-error-prod-silence.test.js): a component whose render() throws, driven through createRequestHandler with NODE_ENV unset. PROD (dev:false) emits NO message (still isolates, siblings render); DEV (dev:true) surfaces it via the isolation template (data-webjs-error); a context-free renderToString falls back to NODE_ENV. Counterfactual verified: neutering the dev-gating makes the PROD case leak.
• Full node suite 2397 pass; core render/suspense + server dev + test/ssr 493 pass (caught and fixed a regression where loadingTemplates's local dev param was wrongly referenced as opts.dev).

Definition of done

• Tests: integration above. Browser/e2e: N/A because the change is server-side SSR error rendering; the observable difference is the ABSENCE of a leaked message in prod, asserted by the integration test. The blog e2e (82) and the streaming/suspense suites cover the happy path, which is byte-identical.
• Docs: N/A because the user-facing docs (error-handling, components.md) say "dev"/"prod", not "NODE_ENV", so they were aspirationally correct; this fix makes the "silent in prod" claim actually true under webjs start. The deployment doc already documents that webjs keys on the dev flag.
• MCP / editor plugins: N/A because no introspection surface, rule, or grammar changed.
• Dogfood: blog e2e 82 pass; website / docs / ui-website boot 200 in dist mode, no broken preloads. (Note: @webjsdev/core dist is gitignored and CI rebuilds it from src; I rebuilt it locally so the createRequestHandler path exercises the fix.)
• Version bump: core + server are owed a follow-up patch at the next release; not in this PR.

[vivek7405]

Went through the threading end to end. The fix is right and the back-fill of ctx.dev is safe (request-scoped ctx, only set when undefined). Two things on the record. First, my original integration test passed for the wrong reason: under createRequestHandler the render effectively sees NODE_ENV=production, so reverting the dev-gating did not make the PROD case leak, only the DEV case flipped. The real counterfactual lives at the renderToString({suspenseCtx:{dev}}) layer with NODE_ENV deleted, which I added. Second, processSuspenseElements was not getting the dev flag, so a throw inside an inline boundary on a ctx-absent render fell back to NODE_ENV (masked by a double-render). Both fixed.

[vivek7405]

Second look at the threading completeness. Caught that my earlier processSuspenseElements fix never actually landed: a counterfactual git checkout reverted it before I committed, so the helper was still re-deriving ctx && ctx.dev instead of taking the dev param. That left a throwing component inside an inline (ctx-absent) falling back to NODE_ENV (masked only by the outer walk re-rendering it, which also wastes a render). Re-applied it and unified BOTH suspense helpers (processSuspenseElements + streamSuspenseBoundaries) onto the threaded dev param, so they can no longer diverge from the walk.

[vivek7405]

Last pass: the dev threading is complete across all four error paths (the walk, an inline non-streaming boundary, a streamed boundary, and a rejected streamed boundary), and the only remaining ctx.dev reads are the legitimate back-fills. Two JSDoc blocks (injectDSD + streamSuspenseBoundaries) had not documented the new dev param; filled those in.

[vivek7405]

Final pass is clean. The dev flag threads through every SSR error path now (the walk, inline and streamed boundaries, and a rejected streamed boundary), the back-fill is the only ctx.dev read and is request-scoped, and the core-layer counterfactual genuinely fails if the gating is removed. Ready to merge.