fix(sse): widen max_tokens/tool_use exemption to lone empty text blocks (#12968)

[diegosouzapw]

Closes #12968

Root cause (short)

detectMalformedNonStream in open-sse/utils/diagnostics.ts only exempted a terminal stop_reason of max_tokens/tool_use (the #9971 fix) when content was a completely empty array (content: []). A tiny max_tokens probe against an Anthropic-compatible provider shim instead returns content: [{type:"text", text:""}] — one block with no visible text — which took a different path through the function and fell straight to empty_choices regardless of stop_reason, producing a false upstream_empty_response → 502. errorClassifier.ts's raw-body isEmptyContentResponse already handles this shape correctly (it short-circuits on content.length > 0), so the mismatch was only in the post-translation classifier.

Fix

Hoisted the stop_reason exemption check (max_tokens/tool_use) out of the content.length === 0 conditional so it applies whenever detectMalformedNonStream finds no visible output — regardless of whether content is [] or holds only textless/empty blocks. No change to errorClassifier.ts (already correct for this shape). The exemption's values (max_tokens, tool_use) are unchanged, only its scope test widened.

Regression test

tests/unit/issue-12968-anthropic-shim-empty-text-block.test.ts — built from the exact upstream body captured in the issue.

RED (unfixed code):

✖ #12968 max_tokens probe with content:[{text:''}] must NOT be flagged empty_choices (527ms)
  AssertionError [ERR_ASSERTION]: expected legitimate truncated-probe response to pass through, got reason=empty_choices
  + actual - expected
  + 'empty_choices'
  - null
✔ #12968 control — content:[] + max_tokens already exempted (#9971)
✔ #12968 control — empty text block with end_turn stop_reason stays flagged
tests 3, pass 2, fail 1

GREEN (fixed code):

✔ #12968 max_tokens probe with content:[{text:''}] must NOT be flagged empty_choices
✔ #12968 control — content:[] + max_tokens already exempted (#9971)
✔ #12968 control — empty text block with end_turn stop_reason stays flagged
tests 3, pass 3, fail 0

Gates run

• npx eslint --suppressions-location config/quality/eslint-suppressions.json open-sse/utils/diagnostics.ts tests/unit/issue-12968-anthropic-shim-empty-text-block.test.ts → clean
• npm run check:open-sse-typecheck → exit 0
• node scripts/check/check-file-size.mjs → no violation on touched files (pre-existing open-sse/utils/stream.ts drift is base, not mine — identical line count on origin/release/v3.8.51)
• node scripts/check/check-complexity.mjs → clean, no violations on touched files
• node scripts/check/check-cognitive-complexity.mjs → OK, 1276 violations vs baseline 1437 (no regression)
• node scripts/check/check-test-discovery.mjs → OK, new test file discovered
• Existing suites re-run: tests/unit/diagnostics.test.ts, tests/unit/issue-9971-empty-choices-contentless-claude.test.ts, tests/unit/diagnostics-claude-thinking-5108.test.ts → 50/50 pass, no regressions

Existing tests aligned

None — no existing assertion encoded the old buggy contract; the new exemption scope is strictly additive to an already-correct set of cases (verified against all pre-existing Claude-shape tests in the files above).