Batch pre-export terminal gates into one report per run

[MaxGhenis]

Why

Build M attempts 9, 10, and 11 each burned a ~2h release run to surface ONE failing pre-export gate group serially: zero-support targets, then export-mass parity, then the reform-coverage smoke. The calibration battery, input-coverage, and export-mass-parity gates all evaluate before the H5 write, so a single failed run can report every failing pre-export group at once — 3 diagnose cycles → 1.

What

• Batched pre-export report in tools/build_us_fiscal_refresh_release.py: the calibration battery collects into terminal_gate_failures instead of raising; the input-coverage and export-mass-parity evaluations run under degraded-mode guards (an evaluation crash with earlier failures on record becomes one more failure line instead of masking them; a crash on a clean run propagates unchanged); ONE batched raise fires just before the H5 write. Green-path behavior is unchanged, and certification manifests are never written for a gate-failed artifact.
• Kept original raises for the reform-coverage smoke and the take-up contract: each already evaluates and reports its full failure set internally, and both require the written H5 that a gate-failed run must not produce.
• expected_sign: "either" support in reform_coverage_smoke.py (scores abs(effect) against the floor, fails with "an effect in either direction") and in the release_input_coverage.py validator. The farm probe (qbi_farm_operations_income_exclusion) moves to "either" with a two-channel rationale: the measured ASEC FRSE leg is net-positive while the donor-pinned PUF Schedule F leg is net-negative per Pin signed sparse QRF outputs' net mass to the donor instrument (farm sign flip; #432) #435 — the net direction is a frame property, not an input-coverage property. Manifest regenerated with the generator (158 required / 8 exclusions / 37 probes).
• Tests: test_either_sign_probe_passes_both_directions_and_keeps_floor_teeth, test_probe_rejects_unknown_expected_sign; the farm case in test_shipped_farm_probes_each_remove_only_the_bound_qbi_leaf updated to "either"; test_main_writes_diagnostics_before_post_calibration_gate_failure now asserts the batched pre-write semantics (diagnostics written + RuntimeError leading with the battery failure text).

Gate-integrity notes

• No gate is weakened: every group that raised before still aborts the run before any certification manifest is written; the only change is that a failed run reports the complete pre-export failure list instead of the first group.
• The "either" farm probe keeps the magnitude floor, so a dead farm channel still fails the probe.

🤖 Generated with Claude Code