Active-reasoning veto misses resumed reasoning outside the {thoughts, code} content_type allowlist (fail-open)

[StartupBros]

evaluateProTurnCompletion's active-reasoning veto only treats a node as "still reasoning" when both conditions hold:

(message.content?.content_type === "thoughts" ||
  message.content?.content_type === "code") &&
message.metadata?.reasoning_status === "is_reasoning"

(src/pro-final.ts, current main at 12ed925a5.)

If a resumed-reasoning node carries reasoning_status: "is_reasoning" but a content_type other than thoughts/code, the veto doesn't see it. A terminal-shaped interim text produced before that node then passes every other check (reasoning_ended precedes it, it is the only terminal leaf) and gets returned as {done: true, ...} — a false-positive terminal on a mid-reasoning progress text.

This reproduces the exact failure the function's design comment says it exists to prevent: "Pro can emit several short progress texts carrying both flags and then resume reasoning."

Reproduction

A mapping with:

recap(reasoning_ended)
  -> terminal-shaped interim text
    -> resumed reasoning {
         content_type: "reasoning",
         reasoning_status: "is_reasoning"
       }

Current result (executed against a logic-identical plain-Node port of pro-final.ts):

{
  "done": true,
  "finalText": "Quick take: looks fine at a glance.",
  "finalMessageId": "msg-interim",
  "modelSlug": "gpt-5-6-pro",
  "finishReason": "stop"
}

Expected: done: false.

Suggested fix

const activeReasoning = turnMessages.filter(
  ({ message }) => message.metadata?.reasoning_status === "is_reasoning",
);

That is deliberately more conservative: reasoning_status is the state signal; content_type names the node kind. Dropping the kind allowlist prevents new/renamed reasoning node types from failing open. It may reject additional ambiguous mappings (safe false negatives), but it cannot newly accept an in-progress turn.

I reproduced this red/green in a Vitest regression: the test fails before the one-line change (expected true to be false) and passes after. Happy to open a small PR with the fix and regression test if useful.