fix: correctness quick-wins (batch A) — prod WASM opt, error surfacin…

[diegokingston]

…g, a11y, type/report bugs

Six independent fixes, all present on current main:

• deploy-gh-pages.yml: drop --no-opt so the PRODUCTION WASM solver runs wasm-opt (binaryen) — the shipped binary was unoptimized (~15-40% slower/larger). CI keeps --no-opt for build speed.
• Toolbar.svelte (2D + 3D): solveCombinations{,3D} returning a string error (e.g. a load case is a mechanism) was silently dropped while the success toast fired anyway. Now surface the error.
• WhatIfPanel.svelte: four catch {/ignore/} blocks + silent string-error returns hid every singular/instability during what-if exploration, leaving stale results with no feedback. Capture and show it inline (a toast per 60ms re-solve tick would spam); clear on close.
• App.svelte: toast container had no live-region semantics. Each toast now announces to screen readers — errors assertive (role=alert), rest polite (role=status), aria-atomic.
• auto-verify.ts / verification-service.ts: AutoVerifyModelData declared its section/material maps too narrowly (no a/iz/iy/j/e), so the steel path read material.e / section.a etc. as TS2339 (runtime was fine — callers pass full objects). Widen the type to the optional fields actually read, add a presence guard in the steel path (narrows the types + skips malformed input), and drop three as any casts (fu/tw/tf).
• calc-report.ts: the element table read e.hingeStart/hingeEnd (removed in the typed-Release migration) → the hinge column was always "—". Use releaseI/releaseJ.mz.

Verified: vite build passes; auto-verify + calc-report suites green; the #2 TS2339s are resolved with no new errors. (Pre-existing, left as-is: harmless d.uy/d.rz legacy-field fallbacks in the calc-report 2D table, and some unused imports.)