fix(qa): add EOS stop_tokens to Golden Output gate — closes phantom #1864 cuBLAS

[noahgift]

TL;DR

Closes #1864. The "cuBLAS FP8 7B Q4K gibberish" was a missing-stop_tokens config in the QA gate, not a numerical bug. User instinct ("could this be a chat template or something simple?") was correct.

Falsification path

Hypothesis	Falsifier	Result
Deep cuBLAS FP8 numerical bug	Run my Stage A reproducer	Yes, single-step argmax mismatches by 1057 vs 75311, correlation 0.987
Bug surfaces in user-visible HTTP	apr serve run <7B GGUF> + curl /v1/chat/completions	❌ User-visible path returns '2+2 equals 4.' CORRECTLY
Then why does Golden Output FAIL?	Inspect gate's gen_config	..Default::default() → stop_tokens: Vec::new()
What does apr serve set?	Read cuda_chat_backend.rs:113	stop_tokens: vec![eos_token_id]

Five-whys:

1. Golden Output FAILs with <|im_start|> repeats — why?
2. Generation runs the full 512 max_tokens — why?
3. No EOS in stop_tokens, defaults to empty — why?
4. The gate's gen_config uses ..Default::default() without overriding it — why?
5. The author of the gate didn't realize Default::stop_tokens = Vec::new(). Root cause = config drift between gate and production path.

Live discharge (RTX 4090, this host)

Pre-fix:

$ apr qa qwen2.5-coder-7b-instruct-q4_k_m.gguf
✗ FAIL Golden Output GPU output failed (CPU passed): gibberish (fragment "<|im_start|>" repeats 3+ times)

Post-fix:

$ apr qa qwen2.5-coder-7b-instruct-q4_k_m.gguf
✓ PASS Golden Output 2 golden test cases passed

Diff

Two 4-line changes in crates/apr-cli/src/commands/golden_output.rs:

• CPU path (line 50-57): rename bos → specials, add stop_tokens: vec![specials.eos_id]
• GPU path (line 153-170): same pattern

Total: 16 insertions, 4 deletions in 1 file. The "multi-day cuBLAS expert work" estimate was wrong — this was 5 lines.

What this revokes

• v0.35.0 release HELD justification → UNBLOCKED
• SPEC-CUBLAS-FP8-7B-FIX-001 Stages C-F → no longer needed; Stage A reproducer + Stage B per-layer instrumentation remain useful but not blocking. Spec status → SUPERSEDED.

What still holds

• The reproducer in PR feat(cublas-fp8-7b-stage-a): deterministic reproducer for #1864 (SPEC Stage A) #1884 (Stage A) correctly observes ~3e-3 abs-diff per element between CPU vs cuBLAS FP8 single-step forward. This is expected FP8 precision, not a defect. Stage B (PR feat(cublas-fp8-7b-stage-b): per-layer parity dumps + drift signature (SPEC Stage B) #1887) documented it correctly as "FP8 precision floor".
• PR fix(wgpu-parity): multi-step gate catches 7B Q4K autoregressive drift (#1864 wgpu) #1876 wgpu multi-step parity gate remains a real improvement: catches autoregressive drift that would otherwise produce silent gibberish on wgpu-only paths.

Methodology lesson

Worth saving to memory/feedback_falsify_simple_before_deep.md:

When a test gate FAILs with a complex-looking symptom (e.g. "FP8 numerical drift"), the first falsifier should be: does the user-visible code path that the test purports to verify ALSO fail, or just the test gate? If only the test fails, the bug is in the test config, not the implementation. ~2 minutes of apr serve + curl would have saved days of bisection.

Test plan

• apr qa <7B Q4K GGUF> now PASSes Golden Output (live verified)
• apr qa <1.5B Q4K APR> continues to PASS (no regression)
• apr serve run <7B> + curl produces correct output (verified pre and post)
• cargo test -p aprender-contracts --lib lint_passes_on_real_contracts
• CI: workspace-test, fmt, contracts-lib

🤖 Generated with Claude Code

[noahgift]

Subsumed by #1894 (release PR for v0.35.0). The squash-merge into release/v0.35.0 preserves the per-PR commit message and changes — see PR #1894 commit log. Closing as superseded.