calibration: partition verifier false-positive rates by the new grader-independence fact

[macanderson]

Problem

#1795's fix (see the PR that closes it) stamps LadderSnapshot::verifier_independent
on every model verdict: Some(false) = the verdict call resolved to the worker's
own model. The 46%-agreement measurement cited in
crates/stella-pipeline/src/verify.rs was made under exactly that self-graded
condition, but the calibration fold cannot see the distinction: CalibrationReport
(crates/stella-pipeline/src/replay.rs) tallies verifier passes/false-positives
as one cohort.

What to build

Extend the calibration fold to partition the verifier cohort by
snapshot.verifier_independent:

• verifier_passes_self_graded / _independent / _unknown (pre-Nothing enforces verifier != worker for the verdict call — one configured provider means the worker grades itself with a prose warning #1795
  snapshots and worker-unresolvable runs stay unknown — never assumed).
• The same split for reconciled/false-positive counts, so
  stella verify calibration (or whatever renders render_calibration) can
  answer THE question the field exists for: is a self-graded PASS measurably
  less trustworthy than an independent one?

Where to look

• crates/stella-pipeline/src/replay.rs — CalibrationReport, calibration,
  render_calibration, and replay/late_reconciliation_tests.rs.
• crates/stella-protocol/src/ladder.rs — LadderSnapshot::verifier_independent.
• Turn on the independent checker and measure whether it helps #1284 (measure the independent checker) — this is the reading side of that
  measurement.

Done when

The report renders the two cohorts' false-positive rates side by side (with
unknown excluded, never folded into either), pinned by a fold test over a
scripted stream carrying both polarities.

[macanderson]

Triage 2026-08-06: P2 — a measurement/reporting gap, not a defect: #1795's verifier_independent fact is stamped but the calibration fold folds both cohorts together, so the report cannot yet answer the question the field was added to answer (is a self-graded PASS measurably less trustworthy — was the 46%-agreement number a self-grading artifact?). Plan: as specced; next step is partitioning the CalibrationReport tallies into self-graded / independent / unknown — with unknown excluded from both, never assumed — pinned by a fold test over a scripted stream carrying both polarities, then the side-by-side false-positive render. This is the reading side of #1284.