fix(inference): tokenize endpoint sent wrong field — every token estimate was 0

LlamaCppTokenizer posted {"tokens":[text]} but llama.cpp's /tokenize expects
{"content":text}; the server silently answered {"tokens":[]}, so countTokens
returned 0 for every string. Every context entry carried tokenEstimate=0 and
the whole budget/trim pipeline was blind — live QA saw a 34k-token prompt
sail through a 16384 stage budget (three raw file_read results of plan docs),
crater t/s, degrade output, and fail the stage on validation 3x.

Also guard estimateTokens: a zero count for non-blank content falls back to
the chars/4 heuristic so a lying tokenizer can never blind budgeting again.

infrastructure/inference/llama_cpp/src/main/kotlin/com/correx/infrastructure/inference/llama/cpp/LlamaCppApiModels.kt

@@ -49,7 +49,7 @@ data class Usage(
 )
 
 @Serializable
-data class TokenizeRequest(val tokens: List<String>)
+data class TokenizeRequest(val content: String)
 
 @Serializable
 data class TokenizeResponse(val tokens: List<Int>)

infrastructure/inference/llama_cpp/src/main/kotlin/com/correx/infrastructure/inference/llama/cpp/LlamaCppTokenizer.kt

@@ -19,7 +19,7 @@ class LlamaCppTokenizer(
         httpClient.post("$baseUrl/tokenize") {
             contentType(ContentType.Application.Json)
             accept(ContentType.Application.Json)
-            setBody(TokenizeRequest(tokens = listOf(text)))
+            setBody(TokenizeRequest(content = text))
         }.body<TokenizeResponse>().tokens.map { Token(it) }
 
     override suspend fun countTokens(text: String): Int = tokenize(text).size