comment #302
**STEP 1 field verdict: CONFIRMED (doyle, 2026-09-17 09:05Z). Mechanism located with symbolized stacks.**

**Capture:** one 180 s collector on ENLYZEAM against the release diagnostic build 3142753a (`field-release-3142753a-operator/`, hertz, collector exit 0). 180/180 usable samples, 4 stale canary intervals, peak 4011 ms, every interval starting ~1 s after a `NET_MEET_ROUND` across the three attached subnets.

**What the stacks show (both `spt-broker-net` workers, tids 5852 and 18952, in every interval):**
`ntdll!NtDeviceIoControlFile` ← `NSI!NsiGetAllParameters` ← `iphlpapi!ResolveIpNetEntry2` ← `iphlpapi!SendARP` ← `netdev::os::windows::interface::interfaces`, reached (a) via `iroh::socket::bind` ← `bind_scoped` ← `meet::bind_rendezvous` (one fresh rendezvous endpoint per subnet per 30 s TOTP window) and (b) via `netdev::route::get_default_gateway` ← `netwatch HomeRouter::new` ← `portmapper::ip_and_gateway` (iroh's portmapper actor spawned by that bind). On this box a gateway that does not answer ARP makes each enumeration ~1165 ms; 92 of 248 `NET_IF_ENUM` calls in 180 s were in that class (8 at 2.1 s, 2 at 3.1 s). Synchronous blocked time over 179 s: worker 5852 = 150 s, worker 18952 = 124 s. `NET_MEET_RETIRE` / `NET_CANCEL_MIB` cost ~0 ms (retirement is not the cost).

**Analyzer:** frozen v1 (`6f36563c…`) said INSUFFICIENT_EVIDENCE on all four intervals because its `NET_IF_ENUM` symbol pattern expected `GetAdaptersAddresses`, not `netdev…interfaces`/`SendARP`; every other acceptance rule (both workers captured in-peak, tid-matched sync span ≥250 ms, raw-copy-resume timing) was met. hertz minted v2 (`6314743d…`, DELIVERY `68d442df…`) widening only those two regexes (diff in `raw-copy-resume-v2/regex.diff`); deterministic re-run on the same preserved inputs: summary mechanism **CONFIRMED** (interval 1 with 5 accepted both-boundary captures; intervals 2–4 have one capture each so stay INSUFFICIENT under the unchanged two-capture rule). Forced positive control unchanged byte-for-byte (18/18 pairs, FALSIFIED via boundary-outside). v1 outputs preserved.

**STEP 2 ruling (dispatched to todlando, `fix/302-meet-offload` off main 84e359aa):** (1) run the complete meet-listener lifecycle (bind + serve + retire, and therefore iroh's descendant actors) on a dedicated bounded `spt-meet` runtime, leaving the main endpoint, streams, dials and canary on `spt-broker-net`; (2) bind rendezvous listeners with the portmapper disabled, as Loopback scope already does. Not accepted as fixes: spawn_blocking around the bind, or more `spt-broker-net` workers. Gate: existing pairing/rotation/starvation/attach suites plus a diag-stamped field re-run on ENLYZEAM (0 stale intervals over ≥180 s with all three subnets still meeting). ENLYZEAM stays on the 3142753a diagnostic build until the fix build is staged; ball-b and engine-room are live on it.