Recovery and PR batches, 2026-09-22

Recovery and PR batches, 2026-09-22

The interrupted working tree contained 1,679 modified or new files, including
146 scientific gene re-reviews, 343 publication caches, and 187 history records.
The complete recovered state was committed locally as
2aa3ece27bb2a68b92629358e3eac23d20ff09b8 on
recovery/iba-treegrafter-20260922 before splitting it. A separate local archive
was also made. This recovery publishes completed edits and their evidence; it
does not claim completion of the corpus-wide audit.

Pull requests

Ten initial batches merged independently. The remaining curation batches were
stacked in the order microbial enzymes (#3109), TreeGrafter (#3110), then
autophagy (#3114), with this tracker (#3120) last. Each returns to main after
its parent merges; auto-merge is disabled while its base is another recovery
branch. This preserves one consistent derived GO-GPT report and benchmark
sidecars at every step, including the article's matching Table S8.
Each batch retains the relevant gene history, research, hypothesis reports, and
analytical evidence. Family-review changes travel with the genes they discuss.

Batch Genes PR
evidence-cache 0 #3107
csr1-identity 1 #3108
microbial-enzymes 20 #3109
treegrafter-divergence 13 #3110
dictyostelium 10 #3111
plastid-and-plant 8 #3112
chaperones 12 #3113
autophagy-and-organelles 14 #3114
lipid-metabolism 13 #3115
enzyme-and-rna-specificity 17 #3116
signaling-and-pseudoenzymes 20 #3117
localization-and-divergence 18 #3118
fulltext-flags 5 #3119
audit-tracker 0 #3120

Validation and recovery fixes

Unfinished scientific work

The regenerated progress report records 81 reviewed,
65 awaiting adjudication, and 3281 unreviewed genes
against the immutable 3,427-gene baseline. These are audit statuses, not merge
statuses. reviewed can still carry explicit human follow-up questions.

The recovered local runner records contain 90 report_ready outputs and one
report_recovered_from_cancelled_job output. These are the statuses recorded on
disk, not a fresh assertion about remote jobs. Availability of a report does not
establish that its reasoning has been incorporated. For entries still awaiting
adjudication, inspect the existing report and primary evidence before launching
another request. No new provider jobs were commissioned during PR packaging.

active-work.yaml preserves historical worker assignments and execution handles
from the interrupted session. They are not live coordination state. Consult the
per-gene audit records, notes, hypothesis directories, and
adjudication-requests.yaml for the actual unresolved questions.

The main IBA and TreeGrafter project findings pages are unchanged by this recovery.

Second opinions and follow-ups from PR review

Superseded ACTL8 and Pmp20 donor assessments are preserved in explicitly labeled
comparison artifacts alongside the current gene reviews. They retain the
historical evidence trail without reinstating unsupported conclusions.

Two initial recovery history filenames retain the scaffolder default claude-code
actor slug. Their agent metadata correctly identifies Codex; the immutable
scaffolded session identifiers were retained.

Completion checks

All 1,679 originally recovered paths are assigned to the batches above, except
for the intentionally renamed CSR-1 archive whose bytes and hash are preserved.
The comparison JSON, figure, Table S8 and deterministic benchmark sidecars are
regenerated together. The full audit remains at 81 reviewed, 65 awaiting
adjudication and 3,281 unreviewed; publishing these batches does not change that
scientific-work denominator. Final merge status is recorded on the linked PRs.

Provenance formatting repairs

The tracking PR also repairs 109 inspected separators/word boundaries across 27 recovered gene
reviews and notes, including already merged batches. This does not add scientific
re-reviews or change annotation actions. Parsed source fields, term objects and
verbatim supporting quotations are unchanged; filenames and accession namespaces
are preserved. The exact before/after tokens are recorded in
prose-identifier-repairs.json. Two existing comparator
citations now have fetched publication caches (PMID:7558035 and PMID:9678974).
The NCGR_LOCUS1270 identity fractions retain their original digits and gain readable
spacing. Each changed gene has a matching history record.

The repair map is an explicit, bounded log. It also records the inspected residual
Bcl2 citations and quantitative/figure spacing; it does not claim that every
stylistic spacing choice in the corpus was normalized.