[anurag629]

F-02 task 6 and F-04 task 10: the widget says why a reply failed, instead of showing "Something went wrong" for everything. The maintainer approved raising the size budget for this.

• Before the stream: the client now reads the protocol v1 error body ({ error: { code, retryAfterMs } }) on any non-2xx. A 429 with retryAfterMs shows "Too many messages. Please try again in 42 seconds." and a 503 provider_unavailable shows "The assistant is unavailable right now". Bodies that aren't JSON (0.5 servers, proxy pages) keep the old behavior.
• During the stream: budget_exceeded with scope conversation shows "This chat has reached its limit." Any other scope (IP, or the site's daily total) shows "The assistant has reached its limit for today". A v1 stream that ends without finish now says "The reply was cut off." instead of the network error.
• Five new strings keys (errorRateLimitWait, errorBudget, errorBudgetDaily, errorUnavailable, errorCutOff), so sites can translate them. The i18n page has a table of every error message.
• The wait is shown once, rounded up, not as a live countdown, to keep the bytes down.
• Size budget: the script tag bundle goes from 14 kB to 14.5 kB (now 14.28 kB, up from 13.97) and ESM from 17.5 to 18 kB (now 17.87). CONTRIBUTING says so.
• The vendored protocol type gets scope on the error part, with a test-d check against core.

Checks: build, typecheck, pnpm test:coverage (1152 passed; new client tests for every mapping and for non-JSON refusals, plus widget tests for each visible message and a strings override), lint at the dev baseline, size, publint, website build.