# ARM B — three complete Phase A suite executions, DIAGNOSTIC TREE ed8ecd94 **DIAGNOSTIC TREE `ed8ecd94a8a610815d5c2d6b54d3ed0f9d789d56`, NOT the candidate.** Candidate `a9e786b2` unmoved throughout. Protocol: `.spt/preserved/hertz-294-armB-protocol.txt` (hertz), authorized by doyle. Executed by deployah on hfenduleam, 18:02:28Z–18:58:00Z, 2026-09-10. **OVERLAP UNVERIFIED.** The protocol's named process-lifecycle recorder (`Win32_ProcessStartTrace`/`StopTrace`) failed at SUBSCRIPTION unelevated — "Call cancelled" and "Access denied", the positive control never reached — and nextest 0.9.137 has no message-format file sink, so the non-invasive fallback does not exist here. Sampled presence (`presence.jsonl`, 1765 / 1916 / 1976 records) is preserved as ITS OWN EVIDENCE CLASS with stamped query intervals and (PID, birth, image) identity per KH 7.51 / `REQ-HAZARD-RESUME-CUSTODY-ABA` and 7.58. **No concurrency claim is attached to it anywhere in this report.** ## Conditions, recorded before B1 (RUN.json) Frozen HEAVY verbatim from `golden.yml:158` (1675 B) → `selection.expr` (1743 B); frozen listing **3381 selected tests, subject identities 1, continuation-control identities 0** (positive reads). Fixtures + workspace prebuilt, excluded from observations. Free space 102.9 → 42.9 GiB against the 32 GiB floor, re-checked before every run. Pool claimed for lane `deployah-armB`. CI in-flight 0, re-checked per run. ## Outcomes | | exit | Summary | population | subject | |---|---|---|---|---| | B1 | 100 | `973.295s` — 3380 passed (9 slow, 5 leaky), 1 failed, 2 skipped | 3381 ✓ | **PASS 99.454s** | | B2 | 100 | `1047.917s` — 3379 passed (11 slow, 4 leaky), 2 failed, 2 skipped | 3381 ✓ | **FAIL 49.963s** | | B3 | 100 | `1108.559s` — 3380 passed (10 slow, 5 leaky), 1 failed, 2 skipped | 3381 ✓ | **PASS 92.136s** | Exactly one Summary per run; population reconciles with the frozen listing in all three. ## Subject, keyed by invocation + test + pid BEFORE any exchange join A naive "first `poll.success` in the log" reads a SIBLING's number — two other sync tests run the same instrumented helper in every run. That trap is real and was caught in B1. | run | pid | events | first pull | later pulls | store-init (1st pull) | |---|---|---|---|---|---| | B1 | 7784 | 65 | 182/400, 4.605s | 20/400, 21/400 | 1.869609s | | B2 | 42784 | **28** | **EXPIRY 400/400, 10.165050s** | — | **9.130515s** | | B3 | 2228 | 65 | 306/400, 7.748s | 22/400, 19/400 | 5.739629s | B2's 28 events is a TRUNCATED COUNT ON A TRUNCATED PATH, not data loss: the failing path stops after the continuation, so exchanges 4 and 6 never emit. ## B2 — the reproduction, and the first UNCENSORED observation of this failure `FAIL [ 49.963s] (1986/3381) spt-daemon::sync two_tier_sync_lands_and_gate_refuses_server_side` Subject's own epoch: ``` 38.0929 poll.enter acceptance window opens 38.0931 store_init.enter 47.2236 store_init.exit store-init alone = 9.130515s of a 10.165050s window 47.2378 → 49.4497 four tip lookups, 2.212s, all outcome=missing — straddling the deadline 48.2580 poll.expiry 400/400, acceptance=failed 49.4498 net_stream_send.ENTER SEND-SUBMISSION ENTRY, 1.191840s AFTER the expiry event 49.4506 transport_inbound_registered submission → correlated registration = 764µs 49.4544 continuation observes the row 1.196399s after expiry, 28.803599s of 30s unspent ``` **Correlation established through the DECODED sync id, not `stream_id`:** `transport_request_identity` carries `sync_id: "pull-2", correlated: true`, remote `da8b6a71…`, and the continuation's late row carries the same remote. `stream_id 1` alone is not a unique cross-pull key — QUIC connections reuse wire-stream id 0 (hertz's pre-handoff finding). **What this supports:** in this execution the requester had NOT reached send-submission entry when acceptance expired; the overrun is localized UPSTREAM OF SUBMISSION. The post-write visibility-gap arm is excluded FOR THIS EXECUTION by construction — not because the wire was fast, but because nothing had been submitted by the deadline. **Withdrawn phrasings, in place:** "the first write happened" and "nothing was ever late on the wire" (deployah, then doyle repeating it; caught by hertz and todlando). `net_stream_send.enter` is SEND-SUBMISSION ENTRY, not a measured wire write — KH 7.66(4): an interval is named for what was actually hooked. ## What is NOT claimed - NOT a classification of the golden reds. Signature match (same cell, 400/400, ~10.16s) is not mechanism match; whether 34445961595 and 34481993681 share this locus is UNMEASURED. **Finding 1 remains UNCLASSIFIED as to cause.** - NOT contention, filesystem, or antivirus. Overlap UNVERIFIED; WHY store-init excursed is unmeasured. - NOT a failure rate. Three draws: 182 pass, 400 expiry, 306 pass. No rate, no pooling with Arm A. - NOT a randomized comparison. Confounders STATED, not corrected: A-then-B ordering, B's workspace prebuild and cache changes, sampler overhead. - **Reproduction does NOT unblock the release** — pre-registered in advance precisely so it could not be argued afterwards. ## The one bounded comparison Store-init on the subject's first pull: B 1.869609s / 5.739629s / 9.130515s against Arm A's 15 isolated draws spanning 0.196–3.497s. **Two of the three B values lie outside A's entire observed range.** A statement about observed values only. Store-init is the stage that MOVES and it is the stage that consumed the window in the one expiry. Paired per-run attribution is todlando's; no independent range extrema. ## Separate signature, preserved not filed `spt::daemon_stop_convoy_e2e stop_under_an_api_storm_stays_down_then_start_brings_up_exactly_one` FAILED in **all three** runs: 9.473s, 10.973s, 14.553s. A different signature is not a sync reproduction; nothing is filed against #294. It is a load-reproducible failure at the diagnostic tree and needs its own disposition — doyle's read. ## Also observed, not pooled The two sibling sync tests are separate tests and are NOT comparable to Arm A's subject-only distribution: `concurrent_writes_reconcile…` first pulls 199 / 147 / 166 of 400; `torn_pull_recovers_by_repulling` 116 / 133 / 266 of 400. Suite wall grew monotonically across the three runs — 973.295s, 1047.917s, 1108.559s — recorded as an observation with no mechanism attached. ## AMENDMENT, 2026-09-10 ~19:05Z — an endorsement withdrawn in place This report's author endorsed doyle's inference that the monotonic growth of convoy duration (9.473 / 10.973 / 14.553s) and suite wall (973.295 / 1047.917 / 1108.559s), set against NON-monotonic subject store-init (1.869609 / 9.130515 / 5.739629s, peak in the middle run), ELIMINATES inter-run drift as an explanation for B2's expiry. **That endorsement is WITHDRAWN.** hertz and todlando refuted it independently: 1. Suite wall and convoy duration are OUTCOMES, not a drift meter — "the most-drifted run" is unmeasured, and using two dependent variables as an instrument for a latent one is the same move this arm refused elsewhere (a phase total cannot locate a delay in one cell; a count is not a load). 2. Non-monotonic store-init is compatible with progressive degradation PLUS run-specific variation. It refutes only "drift alone, monotonically", which nobody proposed. **The measured values are unaffected** and stand as reported; only the inference drawn from their shapes is withdrawn. **The elimination that survives is structural, not comparative:** post-submission delay cannot explain B2's expiry, because the expiry occurred BEFORE submission began. **Convoy wording corrected the same way:** the disposition is a TRIAGE decision — nothing filed against #294, not a release blocker, not quotable about `a9e786b2` — with CAUSE UNCLASSIFIED retained. 0-in-CI-at-candidate vs 3-of-3-local-at-diag-tree does NOT exclude an intermittent candidate defect; instrumentation and local conditions are possibilities, not an exhaustive inventory. ## AMENDMENT 2, 2026-09-10 ~19:06Z — carried residue between runs (unpredeclared confounder) hertz asked whether earlier-run children persisted into B2/B3. Measured from the per-run presence samplers (image-path arm), counting processes alive at each run's first sample that were born before that run started, excluding the run's own cargo / cargo-nextest: | run | first sample | carried-in processes | |---|---|---:| | B1 | 18:02:29.717Z | **0** | | B2 | 18:19:38.709Z | **6** (B1's orphans, born 18:03:17–18:04:19Z) | | B3 | 18:38:28.086Z | **12** (B1's six + B2's six, born 18:20:48–18:21:55Z) | **The three runs were therefore NOT the same condition.** At least six orphaned `spt.exe` daemons per suite SURVIVED to the post-B3 census and the next run inherited them. Two scope limits, both hertz's: 0/6/12 are **CARRIED-IN cohorts**, not the total processes present during each run (each run also created its own cohort); and six surviving per run establishes **at least six**, never an exhaustive count of what each run leaked — a child that died before the census leaves no trace here. This violates the protocol clause requiring owned residual test/child processes to be accounted for before the next run starts; it was not done between runs. **No inference is drawn.** B3 carried twice B2's residue and passed — that comparison is NOT taken as an elimination, being the same shape as the withdrawn drift elimination. Carried-residue exposure is a STATED, UNCORRECTED CONFOUNDER: neither implicated nor excluded. **Also withdrawn: "three self-exited."** 18 were censused, 15 confirmed killed by pid, 0 remained at the post-cleanup image-path census (19:03:53Z). The fate of the other three is UNMEASURED. **For any repeat of Arm B:** a between-run residue sweep belongs in the protocol, not a post-hoc census.