Optimization rule doesn't take into account length of tools

[iskhakov]

---

[Shani871]

function estimateToolTokens(tools):
if tools is empty: return 0
payload = stableJsonStringify(tools) # or JSON.stringify
return countTokens(payload) # or approx(payload.length / 4)

function estimateTotalInputTokens(messages, tools):
msgTokens = estimateMessageTokens(messages)
toolTokens = estimateToolTokens(tools)
return msgTokens + toolTokens

In optimizer:

for candidate in candidates:
inputTokens = estimateTotalInputTokens(candidate.messages, candidate.tools)
totalTokens = inputTokens + candidate.maxOutputTokens

if totalTokens > CONTEXT_LIMIT:
    candidate.invalidReason = "exceeds_context_limit_with_tools"
    continue

candidate.score = baseScore(candidate) + TOOL_LEN_WEIGHT * estimateToolTokens(candidate.tools)

[jyoti369]

Opened a fix in #3733 -- getOptimizedModel now counts tool definition tokens (names, descriptions, input schemas) alongside message tokens before evaluating maxLength rules. Uses the same chars/4 approximation as the base tokenizer. Added 5 unit tests.