## Second correction — the first narrowing was still too strong

Follows [5620384403](https://github.com/BigscreenVR/spt-bs-releases/issues/294#issuecomment-5620384403)
and [5620401744](https://github.com/BigscreenVR/spt-bs-releases/issues/294#issuecomment-5620401744).
Raised by todlando; filed by deployah. **All measurements in the first comment continue to stand.**
This withdraws two more *classification* sentences and adds no new claim.

### Withdrawn from the correction comment

> **Boundary defect — MEASURED.**

> …so the boundary is not separated from the behaviour it is testing, and **some fraction of reds
> must be boundary jitter.**

**Both are WITHDRAWN.**

**Proximity to a threshold does not establish that the threshold is wrong.** Two successes observed
near the budget show that successes have been observed near the budget — nothing about the
acceptance policy. Calling that a "boundary defect, measured" is a policy judgement wearing a
measurement's clothes.

**And no fraction of reds can be attributed to jitter.** That attribution requires the *failures* to
be draws from the same distribution as the successes — which is exactly what the right-censoring
prevents establishing. Two successes are also two samples.

### A naming error rode along, and it mattered

The first comment called the quantity **"publication latency."** That was itself an attribution.
Verified at `a9e786b2`, the wait spans:

- requester store initialization (fresh bare Git store on a first pull) — overlapping the poll,
  `tests/sync.rs:188–201` starts the requester thread and then polls;
- **four sequential `git rev-parse` subprocess spawns** — `src/sync.rs` runs the branch-tip loop
  between `net_stream_subscribe` and the first `net_stream_send`; `branchstore.rs:315` `tip()` calls
  `run_git(["-C", …, "rev-parse", "--verify", "--quiet", …])`, one process per ref;
- scheduling, transport, inbound registration, observation — `nethost.rs:1071–1089` documents QUIC's
  lazy visibility: *"the peer's open becomes visible here on its first write."* Opening alone does
  not publish the inbound row.

Only the last stages are "publication." **No stage's cost is measured.** Git preparation is a
concrete candidate mechanism, not a measured cause.

### The complete surviving record

**Measured:**

- successful observed **wait durations**: 4.448 s (176/400, ipc 0.53 %) and 9.940 s (392/400,
  ipc 0.48 %), against a 400-iteration budget whose nominal value is 10.0 s;
- failed executions: **right-censored** at 400/400;
- the same cell on Linux in the same run: 1.523 s **whole-test wall** — a different quantity, which
  cannot supply a platform-ratio denominator against a wait duration;
- the budget is an **iteration count, not a wall-clock deadline**: 400 × (probe + 25 ms), unbounded
  above.

**Unmeasured:** every stage cost above; Defender's contribution to these executions; whether any
eligible stream appeared after expiry.

**Not established:** that the acceptance threshold is wrong; that any red is boundary jitter; the
failure mechanism (delayed vs. absent vs. filtered vs. preparation-bound).

**Finding 1: mechanism CENSORED, acceptance policy UNRESOLVED.**

### Next discriminating measurement (todlando's, recorded — not dispatched here)

Same-exchange timestamps for: requester-thread entry, store-ready, open/subscribe completion, each
tip lookup, first-write submission, inbound registration, first observed eligible row. Keep the poll
unchanged; authorize bounded observation *after* expiry separately if eventual arrival is the
question.

### Unaffected

Gate verdict untouched: **RED, genuine, not void**, 8 of 9 green. `34445961595` remains the verdict
of record; `34481993681` remains a second red at the same cell. Release stays **HELD** — v0.69.0 not
cleared, core#218 unmerged, no tag, no publish, no board close-out. No state-label changes.