cbm-universe status

Verdict: Early and mostly on-plan, but the work isn't finished. The A1 fetch is slow, and the B1 work and its progress entry are uncommitted.

Good

Commit 690026e (grab-ipr keep_pdbs) is small and has a unit test. The README and progress entry are in the same commit, and it is on hermes.
It found the real cause of the incomplete PF00734 set: the fetch state only recorded shard completion, so a plain rerun reported 0 remaining.
B1 downloads three PDB inputs (1CBH, 1EXG, 1NBC) and writes sha256 digests into docs/data-sources.md. git status shows no data staged.
The grab-ipr and cazy-cbm package tests pass (42 passed, 1 skipped; 1 passed).

Concerns

B1 is unfinished. scripts/, docs/results/gate0-inputs.md and the docs/progress.md and docs/data-sources.md edits are uncommitted. The progress entry for t_f1b33d9b is already written, but the card is still "running". The stash message "partial work from 6 concurrent workers" suggests cards ran in parallel and clobbered each other before a restart.
The fetch is slow. data/interpro/PF00734/AF_PF00734 has 5,646 of 13,427 models, up about 146 in the last hour. The log shows repeated "throttled or server error after retries — left for resume". It was projected at about 9 hours per crawl chunk, so cards A1 and the stage 9-15 re-run may take a day or more.
uv run pytest packages -q from the root fails at collection with ModuleNotFoundError: common in the four cazy-cbm test files. Running pytest packages/cazy-cbm works, so this is a rootdir/sys.path issue. It's probably old, but the AGENTS.md per-package gate hides it.
Gate 0 null set is weak. docs/results/gate0-inputs.md defines it as 50 patches from non-CBM proteins, size-matched to the CBM face. The plan has no matching on hydrophobicity or aromatic content, so the null may be too easy.
Spend: I couldn't read ~/.hermes/logs/budget-guard.log because it's outside the allowed directory, so I can't judge it. Output so far is 1 code commit plus the uncommitted B1 work.

Suggested next

1. Commit B1 once scripts/ is checked, then mark the card done.

2. A card to fix the root-level pytest collection (a pythonpath setting or conftest) so uv run pytest packages works.

3. A card to make the PF00734 fetch resilient: back off on throttling, and avoid re-walking all 15,807 accessions on each pass.

Board

Progress log

Added a regression test that shells out from the repo root to uv run pytest packages/cazy-cbm/tests/test_common.py -q, so the package-suite smoke path now guards the root collection/import setup that uses common.
Verified the root collection path with uv run pytest packages -q (71 passed, 1 skipped) after the new test landed.
Next: none for this card.

Hardened PF00734 fetch resume so keep-pdbs reruns only revisit accessions still missing from the processed log, and download_af_model() now short-circuits to an on-disk AlphaFold model before trying network retries.
Added a throttling regression test that simulates a 429 with Retry-After: 7 followed by a 200, plus a resume test that keeps an already-cached PF00734 model out of the download queue.
Verified with uv run pytest packages/grab-ipr/tests/test_fetch.py -q (2 passed); uv run ruff check packages/grab-ipr/src/grab_ipr/fetch.py packages/grab-ipr/tests/test_fetch.py still reports pre-existing legacy warnings in the untouched parts of fetch.py.

Landed the uncommitted B1/B3 source/docs set: packages/cazy-cbm/src/16_cbm_family_interpro.py, scripts/gate0_registry_noise.py, docs/results/cbm-family-interpro.md, docs/results/gate0.md, plus the packages/cazy-cbm/README.md note pointing to the Gate 0 benchmark.
The family InterPro table covers 105 CBM families; 39 have direct/clan provenance and 66 remain manual. The Gate 0 benchmark says the 0.5 Å face-vs-null row does not hold a stable cross-family registry margin.
Remaining work stays outside git in ignored data//results/ caches; no tracked source/doc files are left from this landing.

Added packages/cazy-cbm/src/17_ted_access.py and packages/cazy-cbm/tests/test_ted_access.py for TED consensus-boundary loading from the public API, with a small overlap matcher against the CBM1 envelope table.
Wrote docs/results/ted-access.md with the access method, the Zenodo bulk release facts (2.9 GB TSV.gz, CC-BY-4.0), and a 20-accession comparison table; the 20-protein test found 19/20 overlap with the local envelope and a 0/20 exact-boundary match rate.
Verified with uv run pytest packages/cazy-cbm -q (46 passed, 1 skipped) and uv run ruff check packages/cazy-cbm/src/17_ted_access.py packages/cazy-cbm/tests/test_ted_access.py.
Next: use the TED loader on the fuller AFDB PF00734 set when the boundary cross-check card lands.

Added scripts/gate0_registry_noise.py to score the three Gate 0 CBM faces against the cellulose Iβ hydrophobic face under Gaussian coordinate noise, with a matched null-patch control built from three non-CBM proteins.
Wrote docs/results/gate0.md plus results/gate0/{synthetic_scores.csv,real_scores.csv,sanity_check.png,pdf,family_scores.png,pdf}; the synthetic 3-point sanity check stays separated through 0.5 Å, but the real CBM-vs-null rows do not share a single family-wide margin at 0.5 Å.
Next: block the card with the gate verdict and numbers for the human go/no-go decision.

Added the new cbm-core workspace package with the shared FASTA, HMMER, structure, and sequence helpers, plus its own tests; cazy-cbm now imports those helpers from cbm_core and keeps only compatibility shims under common/.
Wired cbm-core into the workspace/root dependencies, updated the package READMEs, and verified the split with uv run pytest packages/cbm-core packages/cazy-cbm -q (62 passed, 1 skipped) and uv run ruff check packages/cbm-core.
Next: if any other family-independent helper still lives in cazy-cbm, move it into cbm-core before the next Stage 1 / Track B card lands.

Commits

3h ago	`a942415`	cazy-cbm: add root pytest collection smoke test
3h ago	`ea9d4d1`	cazy-cbm: land B1/B3 docs and scripts
5h ago	`d772c28`	cazy-cbm: add TED boundary access note
7h ago	`e5759e8`	cbm-core: split shared helpers into workspace package
7h ago	`2f9d26d`	substrates: add cellulose ibeta slab builder
7h ago	`bde062c`	cazy-cbm: add cbm family inventory
7h ago	`2f560e7`	Gate 0: add CBM input selection
7h ago	`690026e`	grab-ipr: keep PF00734 alphafold models
7h ago	`38b7b17`	Set up autonomous work: hermes branch, progress log, results notes
7h ago	`2046adf`	Initial layout: uv workspace with grab-ipr and cazy-cbm

Wed 21 Oct	$3.74
Tue 20 Oct	$0.99