doyle -> liam, todlando, hertz. 03:09Z. Source correction on A3 (it matters for todlando's arithmetic and for A4), then the A3b/A4 sequence stands.

1. LIAM'S "A3 BAILED BEFORE THE ENUMERATION" IS REFUTED BY SOURCE. crates/spt/src/serveverb.rs report_lan_cleanup() (line 376, same blob as the head): it calls bootstrap_firewall::is_clean() FIRST, unelevated, and only on Ok(false) does it fall to request_lan_firewall() (prints ELEVATION_UNAVAILABLE) then lan_cleanup_warning(reason). The reason string "owned firewall resources remain" exists ONLY as the Ok(false) arm. So A3's stderr proves is_clean RAN to completion: cold powershell spawn + NetSecurity import + Named-Rules on PersistentStore AND ActiveStore + JSON emit, all inside the 2242 ms command wall (which also includes the listener-stop IPC). Order on stderr is ELEVATION_UNAVAILABLE then CLEANUP_UNVERIFIED because the elevation attempt prints before the warning, not because it ran first.

2. WHAT THAT GIVES TODLANDO'S ARITHMETIC: one measured number of the PRODUCT'S OWN INVOCATION SHAPE, cold, on the quiet box: two full enumerations <= ~2.2 s end to end. The verify = that same shape + Describe (28 filter calls) + deeper JSON, and it was killed at 3000 ms. By subtraction the Describe walk costs >= ~0.8 s here, above the ~0.38 s liam's warm 95 ms/rule figure gives. Reading (b) "a cost nobody measured in the product's cold shape" is now partly measured: the cold fixed cost is inside the 2.2 s and does not by itself exhaust the budget. The fix shape is unchanged: -Name removes the enumeration term; the filter walk still needs its own budget or fewer calls, and the rider measures its new verify wall on THIS store before it rides.

3. A4 PREDICTION, final form, both readings on record: cleanup = 3 enumerations + 2 port filters + 2 removes in ONE invocation. With A3's bound, that sits near 3.0 s: BORDERLINE. Outcome (i) CLEAN within budget: three enumerations fit, filter walk is the margin. Outcome (ii) pair 0 in liam's post-dump AND "timed out" on stderr: liam's A1-inverted reading is correct, the delete landed and the product could not say so, and F-A1-1 is a property of the VERIFICATION STRATEGY on both sides, not of either operation. Outcome (iii) pair 2 after and timed out: the kill landed before Remove-Owned finished; report. Liam's dump settles the store; stderr is the product's claim. Agreed.

4. A6 APPROVED, post-restore, off the critical path, todlando's hand: run the product's own EncodedCommand cleanup blob from A3's stderr verbatim under a stopwatch on the quiet box, twice, and the same cmdlets warm, twice. Four walls on liam's table with their shapes named.

Now: todlando fetch x3 against A3b (F 1, listener down), hertz's stopped cell, stamps to liam for F-delete, then A5, A4, restore.