K=1 rowwise pilot machinery: clone the spine, assign the ladder, measure per-area support (#761)

[juaristi22]

Closes #761 (#495 increment 5, workstream G of #665).

The K=1 rowwise pilot machinery: clone the microcosm-built spine, assign geography through the
ratified OA ladder, and measure per-area support — no solve, no binding, no publication. The epic's
K table was sized against the retired 535,080-row certified compact; the spine that replaced it is
52,846 households, so K re-adjudicates from measured support. This PR lands the measurement
machinery; the licensed runs follow on this branch (two rounds per María's ruling: spine-i now as
shakedown + control, spine-j for the K ruling once #750's licensed phase lands).

What lands

commit	scope
1	the local-areas line renames to uk-local-* and its uk/local Logbook scope is ratified — migration + DECLARED_SCOPES mirror + README move together (María's naming ruling: singular uk/local)
2	the household_id mod 10^8 lineage rule restates for the spine: the pool-only modulus is refused on inputs carrying explicit lineage columns, and the plan/manifest gain an additive explicit lineage block (distinct sources, per-channel splits, flag counts, pre-clone grain)
3	the ladder clone path refuses pre-assigned geography columns — stale *_oa columns cannot ride through beside the ladder's own outputs
4	fail-closed --input-sha256 / --ladder-sha256 pins (calibration-seam convention: refuse before any bytes parse or side effect), with an informational cross-check against the committed local_area_crosswalk.json ladder pin
5	expected_uk_ladder_area_support — the exact analytic expectation (constituency household-count share within region × input region mix × K; OA population shares for LA), sharing the sampler's own stage-weight helpers so it cannot drift from the draw
6	per-area diagnostics lift off the solve matrix: uk_area_support_summary (one groupby, numerically identical), matrix-free uk_ladder_area_support_summary, uk_region_mix; ladder manifest + dry-run gain area_support / region_mix blocks and the real build writes area_support_summary.csv
7	--candidate-clone-counts plans rows/bytes/support at every candidate K in one dry-run (dry-run-only surface, refused on a real build); every candidate K runs the identical extracted clone/assign/gate pipeline
8	the driver's docstring names the spine as the sanctioned input — never the certified compact — plus the changelog fragment

What does not change

No solve and no binding: build_uk_rowwise_candidate.py untouched beyond the two-line pipeline
rename, local_doctrine.py byte-identical, fences and adjudication registers untouched. The
Logbook rung stays f100 — rungs are sample-fraction tokens (#624 grammar) and the pilot consumes
100% of the spine at any K. Dry-runs still append no Logbook row. Weight fences (kind carriage,
mass-log currency, exact conservation at 1e-9, the single clone mass record, the write-path
re-gate) already did what #761 items 3–4 require and are untouched.

Still to land on this branch

• Round A (spine-i, 0cce0399…) LANDED: the multi-K dry-run plan and the real K=1 build
  both ran on the licensed machine — ladder gate passed 650/650 constituencies and 361/361 LAs,
  zero missing-geography rows, mass conserved exactly (abs_delta 0.0), IMPORTANCE kind carried
  with one appended mass record, 16,288 distinct sources via the explicit lineage columns, 17 s /
  1.51 GB peak RSS. Constituency min-ESS by K: 15.3 (K=1) → 28.8 → 86.4 → 245 (K=10). The genesis
  uk/local Logbook row is archived and the chain validates. Receipts:
  experiments/761-uk-rowwise-pilot-receipts.md (+ per-K stats JSON); recommendation there is
  K=4 for First calibrated rowwise UK local candidate (#495 increment 6) #762, adjudication pending Round B.
• Round B (spine-j): re-run when UK WAS stage: split private pension wealth out of corporate_wealth (uk-data#452 M2) #750's licensed phase re-cuts the acceptance receipt; the K
  ruling cites spine-j with spine-i as control.
• The written K recommendation with its measurement, for María to adjudicate; the Adjudicate the UK spine grain at the ingest/calibration seam #688 comment
  (clone-count-as-declaration + measured pre/post-clone grain) posts from the receipts.

Note for the reviewer: the hosted-store migration (20260829000000_logbook_uk_local_scope.sql)
needs manual application; until then the licensed run's live insert warns and continues — local
spooling and the git export path are unaffected.

Verification

Ruff clean; ci_test_groups.py --verify ok; the full uk group ran green pre-rebase (the two
test_uk_local_area_crosswalk failures were the moved #795 branch, resolved by rebasing onto the
merged main); post-rebase the delta surface (16 files: logbook, rowwise, ladder, local-rowwise,
crosswalk, census, spec-pin suites) is green. PR CI is the authoritative full run.

🤖 Generated with Claude Code

[vahid-ahmadi]

Automated review pass (Claude Code, high effort, diff only — no execution; test hunks skipped). Three findings.

1. uk_runtime/local_rowwise.py:1046 and tools/build_uk_rowwise_dataset.py:1573 — the support stats count zero-weight synthetics as support

_area_support_stats reports min_rows, median_rows and bottom_by_rows from assigned_households, which counts every assigned row — including the 10,000 zero-weight SPI synthetic households and their half-mass CGT twins. uk_area_support_summary computes nonzero_households immediately alongside, and the stats block drops it.

So the headline per-area support figures overstate real coverage, and the two numbers in the same summary disagree about the same area: one whose only rows are zero-weight synthetics reports a positive min_rows while its min_ess is 0. That is the specific failure mode this measurement exists to catch — an area that looks supported and is not — and the bottom_by_rows ranking, which is presumably how a thin area gets noticed, is the part most affected.

Since nonzero_households is already computed, reporting the stats over it (or reporting both, explicitly labelled) seems the small fix. Worth deciding what "support" means for the synthetic channel generally: those rows exist to shape the taxpayer population, not to carry area mass, so counting them as area support looks wrong in every consumer of this figure rather than just this one.

2. tools/build_uk_rowwise_dataset.py:422 — a pinned build can proceed unverified

--ladder-sha256 is only refused when --crosswalk is explicitly given. On the crosswalk-generated route (args.crosswalk is None, args.ladder is None) the flag is silently ignored and _verify_artifact_pin is never reached.

So passing a pin on that route does not verify anything and does not complain — the build proceeds as though unpinned. A caller who supplied a digest believes they got verification they did not get, which is worse than not offering the flag on that route at all. Refusing the flag as unsupported there, or wiring the verification through, both close it; silently ignoring a pin is the one option that misleads.

3. tools/build_uk_rowwise_dataset.py:1216,1288 — discarded per-area frames in both dry-run paths

_ladder_area_support_diagnostics is called in both dry-run paths with its second return value bound to _area_support_long and discarded, after the concat of both long frames has been built — full per-area frames for ~650 constituencies and ~360 local authorities, once per candidate K. The work is done and thrown away. Either consume it or stop building it.

---

1 is the one I would fix before merge: it is a measurement that reports coverage an area does not have, in the machinery whose purpose is deciding whether per-area support is adequate.