UK rebuild: sequence calibration after the amounts imputations; staging seam has no CGT target parity evidence #704

Design issue from the first gated staging runs with the CGT family (#552 arc; runs of 2026-08-14, main b420d4e).

The finding

The staging seam runs the CGT amounts redraw after calibration: the certified candidate's weights were calibrated (June release, 149 targets including the Table 1 CGT aggregates) against the pre-repair amounts, and hmrc_cgt_gains then changes the amounts without recalibration. Two measured consequences:

1. The redraw moves total above-AEA gains from £94.4bn to £57.7bn (published: £62.9bn) — the calibrated aggregate is no longer anything the staged artifact was optimised toward. The 92% achieved-vs-published level is exactly the "hold levels?" question the UK capital gains: imputed distribution has no top tail, missing 55% of CGT #552 fence deliberately reserved.
2. uk_target_fit reported status: evidence_absent ("missing evidence: parity_evidence") at terminal — the seam produces no parity evidence, so the aggregate consistency is not merely unvalidated, it is invisible to the battery.

The before-column shape also explains why the pre-repair base overshot: calibration hit the gains aggregate on a tail-less amounts column, which inflates mid-band mass — "too big, too flat" is the fingerprint of calibrating levels onto the wrong shape.

Proposed design (for the #665 rebuild)

Order the pipeline so calibration sees the repaired shape: amounts imputation (SPI, CGT) → calibration → terminal battery. Then:

• the Table 1 aggregates are calibrated against amounts that actually carry a top tail, closing the £5m+ residual through weights rather than invented people;
• uk_target_fit has real parity evidence at terminal, and evidence_absent can harden to a failure for release candidates;
• the UK capital gains: imputed distribution has no top tail, missing 55% of CGT #552 fence's promotion path (band-fact target profile) becomes adjudicable on its own evidence, since shape and level are no longer entangled.

Interim, for the current seam: record CGT aggregate parity (achieved vs the two Table 1 targets) as build-record diagnostics so staged artifacts carry the number even while the gate stays evidence-tolerant.

Refs: #552 (fence adjudication), #630 (the unrelated weight-ratio block on the same runs), #665 (migration roadmap this sequencing belongs to), uk-cgt-paper#6 (downstream consumer waiting on a certified artifact).