# ARM A — 15 repeats, no competing test workload, diagnostic tree ed8ecd94

**DIAGNOSTIC TREE, NOT THE CANDIDATE.** Subject sha `ed8ecd94a8a610815d5c2d6b54d3ed0f9d789d56` =
candidate `a9e786b2` + the two-commit instrumentation patch. Candidate unmoved throughout.
Ref: `diag/294-sync-stages`. Box: hfenduleam. Run window 15:11:58Z–15:28:29Z, 2026-09-10.

Authorized by doyle ~15:12Z as a CONTROL for Arm B and a first spread on the margin.
**Not a failure-rate estimate.** Not a reproduction attempt of golden conditions.

## Conditions
- Pool: sequential takeover of hertz's released lane, claimed from its own worktree
  (`pool-claim --pool target --label deployah-armA`, exit 0). Build once before the loop, exit 0,
  nothing recompiled (1.10s).
- **"No competing test workload BY DESIGN" — not "unloaded".** These runs establish the absence of
  a concurrent test cohort; they establish nothing about other machine load (todlando's wording,
  adopted).
- A1 is **first-in-series only**. It is NOT a Defender first-touch execution: the test binary
  `sync-7917650da419dae4.exe` has mtime 15:00:02Z and was built AND executed during hertz's
  attempt2 run (started 14:59:45Z), and the pre-loop build recompiled nothing.
- Command per repeat, identical to hertz's attempt2 shape:
  `cargo nextest run -p spt-daemon --test sync -E 'test(=two_tier_sync_lands_and_gate_refuses_server_side)' --test-threads=1 --success-output immediate --no-fail-fast`

## Harvest, and the defect in it that was repaired
The per-run extractor anchored `^IR294`; nextest INDENTS captured stdout, so the live extraction
wrote 0 events per run while the raw logs held all 65. Found independently by deployah (15:16Z) and
doyle (15:17Z). **No data was lost because the raw logs were preserved beside the extraction** —
re-extracted post-hoc with a whitespace-tolerant pattern.
**Positive coverage check, not a non-empty check:** all 15 runs match the expected 20-stage /
65-event profile exactly. An extractor's zero is only reportable beside a positive read of the source
(`grep -c IR294` = 65 on every raw log).

## Outcomes
15/15 PASS · 15/15 exit 0 · 15/15 exactly one Summary · **0 expiries** · 0 LEAK rows.

## First-pull poll — the margin, which is the whole point of this arm
Acceptance budget is 400 observations at 25 ms.

| | iterations | % of budget | elapsed |
|---|---:|---:|---:|
| min | 45 | 11.2% | 1.116s |
| q1 | 88 | 22.0% | |
| median | 153 | 38.2% | 3.862s |
| q3 | 231 | 57.8% | |
| max | **300** | **75.0%** | **7.599s** |

All 15 sorted: 45, 48, 81, 88, 91, 92, 151, 153, 162, 184, 186, 231, 246, 281, 300.
hertz's attempt2 draw (175) sits inside this range. **6.7x spread across draws with no competing
test workload.** Contributing stages, first pull: store-init 0.196–3.497s; four tip lookups combined
0.522–5.545s.

## Whole-cell wall
43.942s – 70.864s (median 55.957s). hertz's attempt2 pass of the same binary was 90.946s.

**NON-COMPARABILITY NOTE, stated rather than inferred from:** the CI red's cell was 55.685s. That is
a TRUNCATED path — it stops at expiry — while every figure above is a completed pass. The two are not
measurements of the same quantity, so their ordering supports no conclusion in either direction, and
none is drawn here.

## What this does and does not establish
- **Establishes:** under this condition the first-pull poll consumed between 11.2% and 75.0% of the
  acceptance budget across 15 complete passes.
- **Does NOT establish:** any failure rate; any cause of the golden reds; that the condition
  resembles the golden job. Zero expiries here is **NOT REPRODUCED under this condition** — these 15
  waits all COMPLETED and are therefore UNCENSORED observations; they simply did not reproduce the
  **WITHDRAWN IN PLACE, 2026-09-10:** this line first read "Zero expiries here is NOT REPRODUCED
  under this condition — a CENSORED observation about the golden condition." The word *censored* was
  wrong (todlando caught it; doyle, whose framing supplied it, endorsed the correction). Arm A's 15
  first-pull polls each ran to SUCCESS at a known iteration count: they are COMPLETE, UNCENSORED
  observations. Censoring is for an observation halted before the event — which is what the golden
  400/400 EXPIRIES at 10.156-10.200s are. The real limits are that this is a SMALL SAMPLE (a zero
  count in 15 draws, which does not license "rare") and a DIFFERENT CONDITION (no competing test
  workload, against a golden job running 3381 tests). NOT REPRODUCED stands; only its justification
  changed.
  golden failure. **No failure-probability bound is stated here.** One would require an explicit
  sampling model and assumptions this arm did not establish — not because 15 observations can never
  inform such a bound (hertz's correction of my earlier absolute phrasing), but because the model is
  absent. The probability of exceeding 400 under this condition is UNMEASURED, not small. "Censored" belongs to the EXPIRED polls, which stop before the event —
  todlando's correction, 2026-09-10, replacing my earlier misuse of the term here.
- Finding 1 remains UNCLASSIFIED as to cause.

Preserved: per-run raw log, exit, Summary count, re-extracted events jsonl, observation order with
UTC stamps (`order.jsonl`), `ARM-A-SUMMARY.json`, this report.