Coverage gate: add effective-mass (weighted-support) semantics — signal on zero-weight rows must not count as restored

The blind spot

The input-coverage gate verifies each required column carries non-default signal — but signal can sit entirely on rows with zero (or negligible) calibrated weight, making the column fiscally inert while the gate passes. Found on the UK branch: 200k SPI-synthetic rows carry gift_aid/charitable_investment_gifts signal at zero calibrated weight (log-space calibration cannot lift a zero), so the 132-column contract passed while the HMRC family was effectively absent.

The fix (UK branch is implementing it first)

Extend coverage semantics with a weighted-support dimension: a required column must carry signal on rows holding effective population mass above a reviewed floor. Columns with presence-but-no-mass get a distinct distributional/effective-weight status rather than counting as restored.

This issue: adopt the same semantics in the US gate

The US manifest/gate (release_input_coverage.py, #368/#369) has the identical latent blind spot — restored families are verified for signal, not for effective mass under the certified weights. Post-launch: port the effective-mass diagnostic from the UK implementation, add per-column effective-mass stats to the manifest generator, and decide reviewed floors. The sparse-selection path makes this non-hypothetical: L0 selection could in principle retain a column's signal only on low-weight records.

Start with the US manifest/gate in release_input_coverage.py and compare it with the effective-mass diagnostic being implemented for the UK branch. Port the diagnostic, add per-column effective-mass statistics to the manifest generator, and document reviewed floors so presence without sufficient population mass does not count as restored.