Stream completed exact-ensemble boundaries - #237

Implements the selected Phase 4 architecture change from discussion #232.

The exact non-seed ensemble recurrence finalizes each four-boundary partition exactly once. This branch sends those completed values directly into candidate selection instead of retaining an unordered-map entry for every boundary. Seeded and heuristic predictors remain buffered. Tracker-enabled exact predictions also retain the established buffered callback path.

Modularity and science:

• filter order, Boltzmann conversion, Zall addition order, recurrence, and CLI/output schema are unchanged
• virtual updateZ dispatch is retained for derived predictors
• a scoped per-instance mode restores itself during exception unwinding
• all four overlap policies, streamed/buffered outputs, trackers, override dispatch, and reuse are covered
• the added mode flag changes PredictorMfeEns object layout, so downstream C++ clients must rebuild; the binary-program CLI is unchanged

Validation at 9184083:

• GCC 14.4 strict C++23 focused ensemble test: 54 assertions
• full API: 4,319 assertions in 36 cases
• all 20 CLI golden cases
• fresh benchmark parent/candidate build and test gates passed
• raw and canonical benchmark output bytes are identical

Same-host exact 80-by-80 ensemble benchmark:

• wall: 3.110 +/- 0.098 s -> 2.347 +/- 0.023 s (1.325x; 24.5% reduction)
• mean peak RSS: 89,184 KiB -> 7,595 KiB (91.5% reduction; about 11.7x lower)

Evidence: private benchmark commit 30730f0d0932d31019d8461087f921c7e5497ea5. The separate architecture decision record will link this PR and the same evidence.