fix(llmcore): record Claude messages-API output_tokens instead of zero

[Kailigithub]

Summary

llmcore._record_usage(usage, api_mode='messages') hard-coded out = 0. As a result:

• STATS['out'] was never updated for Claude (Anthropic Messages API) sessions — the user-facing cost / running-stats display stayed stale.
• The [Output] tokens=... log line never printed for Claude sessions, while the chat_completions and responses branches both emit it.

This PR reads usage['output_tokens'] through the same _i(...) helper the other branches already use (which coerces None and non-int values to 0) and adds the missing log line so the three branches stay symmetric.

Diff

--- a/llmcore.py
+++ b/llmcore.py
@@ -375,8 +375,9 @@ def _record_usage(usage, api_mode):
         if out: print(f"[Output] tokens={out}")
     elif api_mode == 'messages':
         ci, cr, raw_inp = _i(usage.get("cache_creation_input_tokens")), _i(usage.get("cache_read_input_tokens")), _i(usage.get("input_tokens"))
-        inp, cached, out = raw_inp + ci + cr, cr, 0
+        inp, cached, out = raw_inp + ci + cr, cr, _i(usage.get("output_tokens"))
         print(f"[Cache] input={raw_inp} creation={ci} read={cr}")
+        if out: print(f"[Output] tokens={out}")
     else: return
     STATS.update(inp=inp, cached=cached, out=out)

2 lines added, 1 changed. No behavior change on existing inputs except that Claude output_tokens are now counted.

Tests

tests/test_llmcore_record_usage_messages.py covers:

1. messages happy path — STATS['out'] populated, [Output] line emitted.
2. messages with output_tokens key absent — degrades to 0, no log line.
3. messages with output_tokens: null — _i coerces to 0 (consistent with the other branches).
4. chat_completions regression — untouched, still correct.
5. responses regression — untouched, still correct.

The regression tests for the other branches are intentional: _record_usage is shared, and the diff is small enough that a single typo could break all three modes. Keeping a 5-case suite means future refactors that touch one branch can't silently regress the others.

Verification

python tests/test_llmcore_record_usage_messages.py
# PASS test_messages_mode_records_output_tokens
# PASS test_messages_mode_handles_missing_output_tokens
# PASS test_messages_mode_handles_null_output_tokens
# PASS test_chat_completions_mode_still_works
# PASS test_responses_mode_still_works
# ALL PASS

# Same suite against the unfixed llmcore.py (git checkout main -- llmcore.py):
# FAIL test_messages_mode_records_output_tokens: messages API: STATS['out'] not updated from output_tokens; got 0
# PASS test_messages_mode_handles_missing_output_tokens
# PASS test_messages_mode_handles_null_output_tokens
# PASS test_chat_completions_mode_still_works
# PASS test_responses_mode_still_works

This PR is the 'optimize' branch from a saturated-repo 3-bucket cron run; it pairs with #760 (fix(fsapp) #685) for bug-fix coverage on the same tick.