---
name: win-netstream-redrive-flake
description: Recurring Windows-only CI flake — ConPTY netstream brain-restart ring-redrive exactly-once tests
metadata: 
  node_type: memory
  type: project
  originSessionId: eb2b5080-21d0-44ab-b08e-f5d1c9f2092f
---

RECURRING Windows-only CI flake (confirmed by doyle 2026-06-23, 2nd hit this session).

**Tests (spt-daemon, `crates/spt-daemon/tests/netstream.rs:99`):**
- `sender_brain_restart_redrive_is_exactly_once`
- `receiver_brain_restart_is_gapless_and_exactly_once`

**Symptom:** `assertion left == right failed: eof before the expected chunk count — a gap the ring should have covered` (e.g. ring-timing 19 vs 20). Windows `test` job RED; **Linux green on the identical tree**; n1-gates + traceability green.

**Root class:** ConPTY netstream ring-redrive timing — Windows is slower, so the brain-restart redrive race intermittently sees a chunk gap the ring should have covered. `netstream.rs` last touched 2026-06-08 (db9f1bb) — unrelated to whatever change is being CI'd.

**Hits this session:** todlando's v0.14.0 W2 run (`receiver_brain_restart_is_gapless`) + deployah's v0.14.0 release bump run 28016824687 attempt-1 (both tests). BOTH cleared on `gh run rerun <id> --failed`.

**Publisher handling:** NOT a release blocker. On this RED with Linux-green + byte-identical-to-a-green-PR code → it's the flake: `gh run rerun <id> --failed`, then **verify the NEW attempt per-attempt** (rerun can attach to a STALE attempt — check `.attempt` and the fresh job conclusion, per [[v013-forkpty-groundtruth]]). Do NOT tag on the failed attempt.

**Follow-up:** post-v0.14.0 real investigation — genuine fix or a Windows-timing guard. Candidate to batch with F-018. Related: [[v013-forkpty-groundtruth]] (LIVE attach forward-gap on ring-roll-behind, same ConPTY ring family).