ci(performance): make bundle budgets evidence-based regression budgets

[qnbs]

Context\n\nBundle budgets exist for entry, vendor, chunks, and WASM, but the current policy needs a measured artifact baseline and explicit user-impact rationale.\n\n## Acceptance criteria\n\n- Measure current production artifacts in both relevant build modes and record compression semantics.\n- Define per-artifact/category budgets with modest justified headroom and a baseline-delta regression check.\n- Keep exceptional WASM budgets explicit and documented.\n- Add deterministic tests for custom budget tooling and make failures actionable.\n- Do not slash thresholds or exclude artifacts merely to obtain green CI.

[qnbs]

PWA acceptance linkage: #478 establishes the PWA as a measured first-class product surface. Please make this issue's artifact budgets feed user-facing PWA performance evidence (startup/editor-ready, lazy-feature cost, offline precache/update footprint, parse/execute cost) rather than optimizing bundle size in isolation. In particular, do not move required offline assets out of safe caching merely to improve a size metric, and keep large model/ONNX/DuckDB artifacts under an explicit lifecycle rather than accidental precache.

[qnbs]

Architecture/program reconciliation — 2026-08-24

Keep this issue as the single bundle-budget/measurement authority. #478 PWA performance and #487 mobile performance consume these budgets; they should not create competing byte/chunk thresholds merely to optimize one surface.

Budgets remain evidence-based and user-impact-oriented. Lazy loading is appropriate for genuinely cold features, but must not hide required correctness work, break offline behavior, or move expensive initialization into interaction-critical paths. Browser-local AI/model artifacts under #483/#482 are lifecycle-managed artifacts, not ordinary app-shell bundle budget entries.