refactor(server): drop stale prompt text, duplicated routes and dead options

- lib/system-prompts.ts: delete the extended prompt kept only to reach
  the 4096-token cache threshold of two old Claude models; describe the
  current canvas-first UI in the App Features section
- app/api/chat/route.ts: remove the server-side example cache (the
  browser answers examples itself), the ENABLE_HISTORY_XML_REPLACE
  option and its helper, the unread messageMetadata fields; the
  DEBUG_LLM_PAYLOAD blocks become one describeMessages summary
- app/api/log-feedback and log-save share scoreLatestTrace in
  lib/langfuse.ts
- app/api/admin/test-model and validate-model call a shared
  testModelConnection() instead of one route faking a request to the
  other and passing trust through the body
- lib/dynamo-quota-manager.ts reads the failing item from the
  conditional update instead of a second GetItem
- lib/ai-providers.ts: geminiThinkingOptions helper for the google and
  vertexai cases, drop the unused fourth resolveBaseURL parameter
- edge-functions: always stream, shrink messageSchema, drop model
  aliases nothing sends
- delete app/api/chat/xml_guide.md, parseOutputTokenLimit, unused type
  exports and export keywords on file-internal symbols
- env.example: drop variables nothing reads, document the quota,
  validation and analytics variables, mention the apiKeyEnv array form
  (also in docs/*/ai-providers.md)

lib/output-token-limit.ts

@@ -14,7 +14,7 @@ type WrappedModel = ReturnType<typeof wrapLanguageModel>
 export const DEFAULT_MAX_OUTPUT_TOKENS = 64000
 
 /** Ceiling for the user-supplied override, to catch typos like an extra zero. */
-export const MAX_OUTPUT_TOKENS_LIMIT = 200000
+const MAX_OUTPUT_TOKENS_LIMIT = 200000
 
 /**
  * Below this a diagram cannot come out whole, so a retry would just produce
@@ -81,7 +81,7 @@ export function rejectionText(error: unknown): string | null {
  * "lower than N") would reinterpret unrelated failures, and retrying on a bogus
  * number turns a readable error into an empty diagram.
  */
-function readCeiling(text: string): number | null {
+export function readCeiling(text: string): number | null {
     // Combined input+output ceiling: subtract the input the provider counted,
     // plus a small margin because its estimate is approximate.
     const context = text.match(/maximum context length (?:is|of) (\d+)/i)
@@ -104,13 +104,6 @@ function readCeiling(text: string): number | null {
     return output ? Number(output[1]) : null
 }
 
-/** The usable output ceiling named in a rejection, or null. */
-export function parseOutputTokenLimit(error: unknown): number | null {
-    const text = rejectionText(error)
-    const ceiling = text ? readCeiling(text) : null
-    return ceiling === null ? null : usableLimit(ceiling)
-}
-
 /**
  * Thinking budget the provider adds on top of maxOutputTokens. Bedrock and
  * Anthropic send maxOutputTokens + budgetTokens as max_tokens, so a ceiling in