[justinchuby]

A hazard in merged code

PR #372's PTY harness calls openpty(None, None) — a 0x0 window. ratatui's insert_before_no_scrolling_regions infinite-loops at zero rows, so the harness as merged is one type-ahead test away from hanging a CI job, and the next person to extend it would have had no idea why.

The window is now a real 24x80, with the reason at the call site — a bare 24, 80 reads as arbitrary and someone would eventually "simplify" it back. The CR-not-LF gotcha is cross-referenced there too: terminals send \r, and feeding \n means the line never completes, which presents as a hang that looks exactly like swallowed input.

Both artifacts were found the expensive way, so they now live in the code a harness author will read, not only in a decision record.

Confirmed neutral rather than assumed: test 1 passes at both 0x0 and 24x80; test 2 behaves identically at both. Test 2's empty read under WSL is not a window-size effect — it is the harness's own 30s drain_pty_master timeout against a ~48s debug model load over the slow /mnt/c mount. A byte-for-byte replica with a 90s window returns the expected tok22tok22tok20\r\n. Native CI Linux loads fast enough that both pass.

Context: three negatives on the type-ahead report

This branch began as an investigation into run swallowing typed-ahead input. Input is not being lost. Each hypothesis was refuted with evidence rather than argument:

• The cursor-query theory is dead on both platforms. crossterm 0.29's cursor::position() routes user keystrokes to skipped_events and drains them back into the queue; ratatui 0.30's inline viewport never reads stdin outside init/resize. Verified by withholding the DSR reply 250ms and injecting /help into the open window — it survived.
• Windows cannot lose it either. ReadConsoleInputW reads raw records unaffected by ENABLE_LINE_INPUT; SetConsoleMode does not flush the queue; there is no FlushConsoleInputBuffer anywhere. Instrumented GetConsoleMode (0x01f7) and pending INPUT_RECORD counts across prompt boundaries confirm it, including for a partial line with no Enter.
• The timing gap is closed. Re-run against qwen2.5-0.5b streaming ~19s, with /help injected pasted, delayed, and character-by-character at human speed. Type-ahead survived every time.

Raw bytes argue against the erase hypothesis: /help appears at the banner and at the next prompt — never mid-stream. ConPTY does not echo type-ahead during generation, so there is nothing for a repaint to overwrite.

No speculative fix shipped. The remaining untested path is native-conhost immediate echo, which no automated harness on a headless box can produce — plausibly why it is visible in an ordinary session and not here. Fixing an unobserved mechanism would look like it worked and would not.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 81.38%. Comparing base (eb5b5d4) to head (c2f7ca6).
⚠️ Report is 3 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #393      +/-   ##
==========================================
+ Coverage   80.59%   81.38%   +0.78%     
==========================================
  Files         311      311              
  Lines      121012   121100      +88     
  Branches   121012   121100      +88     
==========================================
+ Hits        97530    98555    +1025     
+ Misses      19504    18558     -946     
- Partials     3978     3987       +9     

Flag	Coverage Δ
cli-ort-linux	83.16% <ø> (+0.16%)	⬆️
cli-ort-windows	82.55% <ø> (+0.17%)	⬆️
mlas	77.91% <ø> (ø)
offline	81.35% <ø> (+0.82%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.
see 10 files with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[github-actions]

🔴 Benchmark Regression Detected

Comparison of criterion micro-benchmarks: PR head vs merge-base, measured on the same runner in the same job (base first → PR second).

ℹ️ Absolute times are informational only — they vary with runner load. The % change column is the reliable signal because both sides ran under identical conditions.

Status	Scenario	Base	PR	Change
🔴	matmul/medium_generic_f16_threads=8/32x512x512	31.79 µs	47.53 µs	+49.5%
🔴	gather/small_bf16_threads=1-internal/4096	563.1 ns	833.3 ns	+48.0%
🔴	gather/medium_f32_threads=1-internal/32768	4.71 µs	6.29 µs	+33.5%
⚠️	matmul/large_generic_f16_threads=8/32x1024x1024	96.69 µs	123.43 µs	+27.7%
⚠️	matmul/large_generic_bf16_threads=8/32x1024x1024	2.10 ms	2.63 ms	+24.9%
⚠️	matmul/medium_generic_f16_threads=1/32x512x512	33.96 µs	42.24 µs	+24.4%
⚠️	matmul/large_generic_bf16_threads=1/32x1024x1024	2.35 ms	2.73 ms	+15.8%
⚠️	add/large_f16_threads=1-internal/4194304	46.95 ms	54.14 ms	+15.3%
✅	gather/small_f16_threads=1-internal/4096	553.5 ns	621.7 ns	+12.3%
✅	gather/small_f32_threads=1-internal/4096	801.9 ns	868.1 ns	+8.3%
✅	tokenization/encode_tokens_per_second	400.72 µs	414.88 µs	+3.5%
✅	gather/medium_f16_threads=1-internal/32768	2.85 µs	2.89 µs	+1.2%
✅	sampling_latency/top_p_per_token	1.06 ms	1.08 ms	+1.1%
✅	gather/medium_bf16_threads=1-internal/32768	2.93 µs	2.93 µs	-0.1%
✅	tokenization/decode_tokens_per_second	6.56 ms	6.54 ms	-0.3%
✅	matmul/large_generic_f16_threads=1/32x1024x1024	89.64 µs	88.20 µs	-1.6%
✅	matmul/medium_generic_bf16_threads=1/32x512x512	673.42 µs	656.42 µs	-2.5%
✅	reduce_mean/medium_f32_threads=1-internal/65536	287.95 µs	278.08 µs	-3.4%
✅	reduce_mean/large_f32_threads=1-internal/262144	1.14 ms	1.09 ms	-4.3%
✅	add/large_bf16_threads=1-internal/4194304	49.79 ms	46.49 ms	-6.6%
✅	sampling_latency/min_p_per_token	402.28 µs	375.29 µs	-6.7%
✅	matmul/medium_generic_f32_threads=1/32x512x512	2.75 ms	2.57 ms	-6.9%
✅	matmul/medium_generic_f32_threads=8/32x512x512	1.61 ms	1.49 ms	-7.5%
✅	matmul/small_generic_bf16_threads=8/1x256x256	41.53 µs	38.14 µs	-8.2%
✅	matmul/medium_generic_bf16_threads=8/32x512x512	703.63 µs	645.72 µs	-8.2%
✅	sampling_latency/top_k_per_token	558.74 µs	512.03 µs	-8.4%
✅	add/small_f16_threads=1-internal/1024	15.41 µs	14.10 µs	-8.5%
✅	reduce_mean/small_f32_threads=1-internal/4096	18.42 µs	16.64 µs	-9.6%
✅	add/large_f32_threads=1-internal/4194304	52.20 ms	46.86 ms	-10.2%
✅	matmul/small_generic_f16_threads=8/1x256x256	40.85 µs	36.37 µs	-11.0%
✅	gather/large_f16_threads=1-internal/131072	17.52 µs	15.35 µs	-12.4%
✅	sampling_latency/greedy_per_token	4.05 µs	3.53 µs	-12.8%
✅	matmul/small_generic_f32_threads=8/1x256x256	63.58 µs	55.12 µs	-13.3%
✅	add/small_bf16_threads=1-internal/1024	16.25 µs	13.88 µs	-14.6%
✅	matmul/small_generic_bf16_threads=1/1x256x256	41.19 µs	35.06 µs	-14.9%
🟢	matmul/small_generic_f32_threads=1/1x256x256	47.74 µs	40.45 µs	-15.3%
🟢	kv_cache/alloc_dealloc_pages	45.81 µs	38.50 µs	-16.0%
🟢	add/medium_f16_threads=1-internal/262144	3.35 ms	2.81 ms	-16.2%
🟢	matmul/small_generic_f16_threads=1/1x256x256	40.28 µs	33.54 µs	-16.7%
🟢	grammar_masking/llguidance_compute_mask/32	105.06 µs	87.17 µs	-17.0%
🟢	gather/large_f32_threads=1-internal/131072	42.72 µs	35.21 µs	-17.6%
🟢	add/medium_bf16_threads=1-internal/262144	3.52 ms	2.88 ms	-18.2%
🟢	logit_processing/seven_processor_chain_per_step	1.76 ms	1.41 ms	-20.1%
🟢	matmul/large_generic_f32_threads=1/32x1024x1024	14.67 ms	10.85 ms	-26.1%
🟢	add/small_f32_threads=1-internal/1024	343.5 ns	242.5 ns	-29.4%
🟢	gather/large_bf16_threads=1-internal/131072	16.97 µs	11.07 µs	-34.8%
🟢	matmul/large_generic_f32_threads=8/32x1024x1024	6.81 ms	4.27 ms	-37.3%
🟢	add/medium_f32_threads=1-internal/262144	5.37 ms	2.73 ms	-49.2%

Visual flags: ⚠️ ≥ 15% slower, 🔴 ≥ 30% slower — calibrated against measured runner noise (~27% worst-case on multi-threaded matmul)

Host info

CPU: Apple M1 (Virtual)
Cores: 3
OS: Darwin 25.4.0 arm64
Rust: rustc 1.97.1 (8bab26f4f 2026-07-14)
Load avg: { 3.94 4.08 7.32 }

What this cannot catch

• Regressions in code paths not covered by these benchmarks (e.g., end-to-end decode with a real model)
• Sub-threshold regressions that compound over multiple PRs
• Performance changes that only manifest under GPU execution
• Latency changes in the ORT integration path (these benchmarks exercise the native Rust kernels)