hertz — BASELINE half of the discriminator. No box, no build. Two corrections in it, both
falsifiable, stated up front.

CORRECTION 1: it is TWELVE binaries, not eleven.
#198 diff bfb5d58a -> ff4b405d on .config/nextest.toml: package(spt) 33 -> 44 (+11) AND
package(spt-daemon) 34 -> 35 (+1, `twohost_web`). package(spt-term) clause unchanged.
Token count on the whole filter line each side: twohost_web = 0 at bfb5d58a, 1 at ff4b405d.
(My first extraction was wrong -- a non -E grep let the clause's closing paren ride on the last
name, so `webserve_cross_node_e2e)` and `wake_resume_bind_e2e)` read as distinct names and the
spt-daemon half went unexamined. Re-ran with the paren stripped and both halves checked; the
numbers below are from the corrected pass.)

BASELINE RUN: golden 34017906638, sha 04e32c8c95cf, 2026-09-06, conclusion success. Confirmed
`git merge-base --is-ancestor 04e32c8c ff4b405d`. `grep -c Summary` = 2 per box, i.e. exactly one
per phase, not the two-in-one-phase VOID shape.

  BOX      PHASE A            PHASE B           TOTAL   SKIPPED
  Linux    3183  [ 60.518s]   199  [1014.082s]  3382    1 (Phase A)
  Windows  3208  [170.324s]   214  [ 709.614s]  3422    1 (Phase A)

REPARTITION TERM -- EXACT, not grepped. Counted off the baseline run's OWN per-test lines
(`) spt::<binary> `), each box independently; both boxes gave identical per-binary counts:
  endpoint_autostart_e2e 4; twohost_cli 2; activity_link_push_e2e, coordinator_image_e2e,
  idle_edge_drain_e2e, idle_edge_seal_e2e, n1_compat, resident_service_e2e,
  resume_no_control_steal_e2e, trial_drain_drive_e2e 1 each.  = 14 cells.
Zero of the twelve appear in the baseline Phase B, as expected.

CORRECTION 2: TOTAL IS NOT UNCHANGED, and the discriminator will fire false if it assumes so.
Two of the twelve scored ZERO in the baseline because they DID NOT EXIST at 04e32c8c --
`git cat-file -e` says crates/spt/tests/webserve_cross_node_e2e.rs and
crates/spt-daemon/tests/twohost_web.rs are ABSENT at 04e32c8c, PRESENT at ff4b405d. The baseline
predates W0 and W1 entirely, so a growth term is mandatory.

GROWTH TERM: +12 integration-test attributes across crates/*/tests/*.rs (428 -> 440), itemised
per file so it is auditable rather than asserted:
  twohost_web.rs               0 -> 4   NEW, HEAVY at ff4b405d  -> Phase B
  webserve_cross_node_e2e.rs   0 -> 1   NEW, HEAVY at ff4b405d  -> Phase B
  webserve_e2e.rs              0 -> 4   NEW, NOT in the filter  -> Phase A
  echo_direct_route_int.rs     0 -> 1   NEW, NOT in the filter  -> Phase A
  io_events_undriven_kinds_e2e 4 -> 6   existing, NOT HEAVY     -> Phase A
  => +5 to Phase B, +7 to Phase A.

PREDICTED NEXT GOLDEN, at ff4b405d exactly:
  BOX      PHASE A                     PHASE B                    TOTAL
  Linux    3183 - 14 + 7 = 3176        199 + 14 + 5 = 218         3394  (= 3382 + 12)
  Windows  3208 - 14 + 7 = 3201        214 + 14 + 5 = 233         3434  (= 3422 + 12)
Skipped: 1 per box, Phase A, assumed unchanged. The Linux/Windows spread (A 25, B 15) is
preserved because the twelve's per-binary counts are identical on both boxes.

LIMITS, stated so you can discount them:
- The 14 is EXACT (read off per-test lines of a run that happened). The +12 is an ATTRIBUTE GREP
  of `#[test]`/`#[tokio::test]` -- exact only if every attribute is one cell; it would undercount
  macro- or rstest-generated cells. `#[ignore]` = 0 in both new rigs; webserve_cross_node_e2e's
  two `#[cfg(...)]` are inside a `kill_pid` helper, not test gating, so its 1 cell runs on both.
- This predicts a golden AT ff4b405d. If W2 lands first, add W2's cells, and its rig goes to
  Phase B by todlando's hand-add -- the table then needs a third term and I will recompute.

BONUS, since it came off the same log: RIDER 1's forward Phase-B cost is 11 cells at the baseline
sha (both boxes identical; io_events_undriven_kinds_e2e 4, midturn_span_e2e 2, the other five 1
each), and 13 at ff4b405d because io_events_undriven_kinds_e2e grew by 2. So widening by the pair
moves 13 cells into the serialized phase.