Decouple ACS weeks carriers from unemployment (greens weeks_unemployed incidence)

PROGRESS.md

@@ -1,4 +1,95 @@
-# Battery package 3 progress
+# Battery residual-fixes progress
+
+## Residual FIX packages — 2026-08-22 22:45Z
+
+### State
+
+Complete. Implementation and complete five-shard verification are committed on branch
+`battery-residual-fixes`, stacked on `battery-pkg3-two-part`. The only new
+generating change removes a PUF-only unemployment-compensation condition from
+the late ASEC-to-ACS `weeks_unemployed` carrier calibration. The other named
+checks were either already green under pkg3, outside their adjudicated
+criterion, or traced to an upstream owner rather than changed locally.
+`FINAL_REPORT.md` contains the per-check disposition and exact serial-host
+handoff. The same five package suites, repository-wide Ruff, and whitespace
+checks pass on this documentation-only closeout tree. No pool build,
+gate/band/ceiling/fold/seed tuning, exclusion change, push, certification, or
+publication has occurred. Earlier sections of this journal are historical.
+
+### Done
+
+- Read `CLAUDE.md` and the complete GitNexus debugging skill instructions.
+  The graph query/context tools are unavailable in this session, so direct
+  source, call-site, history, and checkpoint tracing is the documented
+  fallback.
+- Confirmed the tracked worktree was clean at `bb94f789` and that
+  `FINAL_REPORT.md` is the established final-report output file.
+- Attempted the required `uv sync --all-packages --extra us` before any
+  repository edit. The default cache is outside the sandbox; a writable-cache
+  retry then failed because network access is disabled. Copied the complete,
+  lock-compatible pkg3 environment, attempted an offline reconciliation, and
+  repointed its five editable workspace packages to this worktree. Import
+  checks resolve every `microcosm.*` shard here and `quantile-forest==1.4.2`.
+- Preserved the owner-only build boundary: the baseline and pkg3 gate files
+  will be read only, and no pool command has run.
+- Ran `uv run --no-sync pytest` separately for the final startup tree's
+  `microcosm-frame`, `microcosm-fit`, `microcosm-calibrate`,
+  `microcosm-data`, and `microcosm-build` test roots; every command exited
+  zero with only established skips and warnings. Repository-wide
+  `uv run --no-sync ruff check .` and `git diff --check` also pass.
+- Recomputed the mirror-deduplicated battery universe directly from the
+  SHA-pinned gate artifacts: baseline has 127 unique failures and pkg3 has 114;
+  pkg3's `terminal_gates` and `agreement_gate` are byte-equivalent JSON.
+- Confirmed the two adjudicated adult-care expense checks reached green
+  incidence (`1.024210809`) and insufficient 1% QED support after pkg3's
+  qualifying mutable-carrier selection and mutable-positive amount map. The
+  separately red incapacity flag is not one of those two expense checks.
+- Confirmed unemployment-compensation's assigned positive-amount QED is
+  exactly `0.0`; its sparse carrier-incidence failure is a distinct unassigned
+  criterion. Confirmed Schedule D's assigned joint-parent incidence is green
+  at `1.092661186`; its new QED failure follows still-red PUF-tax parent amount
+  shapes and remains with that lane.
+- Diagnosed SSI without changing a gate or formula: take-up is `true` on every
+  pkg3 row, eligibility is slightly lower on ACS, and the excess appears only
+  after the upstream countable-income test. No SSI-local calibrated surface is
+  authorized by the adjudication.
+- Removed the late weeks/UC coupling from the closed policy, production owner,
+  receipt validators, fixtures, and authored lineage identity. Added an owner-
+  level regression proving zero-UC ACS recipients remain valid mutable weeks
+  carriers while integer donor support and immutable bytes are preserved.
+- Extended the SHA-pinned pkg3 checkpoint replay with historical and current
+  carrier scopes. The current code admits all 34,293 mutable ACS rows rather
+  than only 32 positive-UC rows, produces 2,174 positive rows on integer donor
+  support `1..46`, is not capacity-limited, validates exactly, and predicts an
+  incidence ratio of `1.000668542` rather than `0.031371146`.
+- Ran the focused lineage, kernel, stacked-owner, H5, and publication-validator
+  matrix after the fix; all selected tests and touched-file Ruff checks pass.
+- Ran `uv run --no-sync pytest -q` separately for the complete implementation
+  tree's `microcosm-frame`, `microcosm-fit`, `microcosm-calibrate`,
+  `microcosm-data`, and `microcosm-build` test roots. Every shard exited zero;
+  the build shard reached 100% with one expected skip and only established
+  warnings. Repository-wide `uv run --no-sync ruff check .` and
+  `git diff --check` also pass.
+- Audited host checkpoint reuse. Although pkg3's `assembled` stage is the
+  semantic pre-change cut, the changed policy is hashed into its checkpoint
+  identity and the CLI has no safe stage-import override. `_LANE-NOTES.md`
+  therefore records an exact cold 1% command in a fresh residual namespace;
+  copying or rebinding the pkg3 manifest is explicitly forbidden.
+- Committed the generating repair, regressions, pinned replay, adjudication,
+  handoff, and green executable verification as `eaba1eab`.
+- Replaced the stale pkg3 report with the residual lane's complete per-check
+  `FINAL_REPORT.md` and preserved the live-tree-safe split host path variables.
+- Reran all five package test roots, repository-wide Ruff, the live-tree
+  retired-data-package guard, and whitespace checks on the final documentation
+  tree; every command exited zero with only established skips and warnings.
+
+### Next
+
+1. The serial host owner queues the exact guarded 1% rerun recorded in
+   `FINAL_REPORT.md` and `_LANE-NOTES.md`; this lane does not run it.
+2. The owner recomputes the mirror-deduplicated residual gates and adjudicates
+   that new host evidence. Certification and publication remain outside this
+   lane.
 
 ## Main/F0 calibration-policy merge — 2026-08-22 02:09Z