NEXT_TASKS: reconcile to 2026-07-30 (PR #268, issues #269-#273, PR #274)

[realmarcin]

Backlog maintenance only — no code or data changes.

Records the network-rendering batch that began from the one remaining actionable in-repo item on the web design review (#199), plus the CI hygiene work it turned up. The other four long-standing open issues (#259, #183, #182, #30) are all blocked on things this repo cannot supply — publisher access, curator judgment, sibling-repo schema — which the new section states explicitly so the next reconcile doesn't re-litigate it.

New section, Per-community network rendering + CI hygiene (2026-07-30):

• PR Render only the interaction types a community actually has (#199) #268 (merged) — the legend now renders only the types a community actually has (857 of 3,000 legend rows were meaningful), the palette is CVD-verified with the full-set minimum going from ΔE 0.0 to 15.4, and tests/test_network_palette.py gates both enum coverage and the ΔE floor. Closes Network interaction palette is not colorblind-safe; COMPETITION and PREDATION are near-identical even in normal vision #269 and No gate keeps the network palette in sync with InteractionTypeEnum #271.
• Interaction type in the network diagram is encoded by color alone; nine categories exceed what color can carry #270 (open) — interaction type is carried by colour alone, and no nine-colour set is safe under all three deficiency types, so a redundant channel is the real fix rather than further palette tuning. Notes that the UMAP already solves this with a symbol scale.
• PR Repair network-quality.yml so it parses, and make the LLM job manual-only (#272) #274 (open) / network-quality.yml is invalid YAML and has failed every run for at least 15 runs #272 — network-quality.yml could not be parsed and had never run. Also records the two conditions that could never have fired even after it parsed, since that is the kind of thing a future reader would otherwise have to rediscover.
• Triage the 26 network audit findings, fix audit-network --json, then restore the hard gate #273 (open) — the 26 audit findings split into 22 soft DISCONNECTED and 4 genuine dangling references, plus the auditor's missing severity levels and the broken audit-network --json.

Two notes aimed squarely at whoever picks this up cold: the surviving src/communitymech/templates/community.html.j2 is live and must not be confused with the stale top-level duplicate #199 asked to delete, and the taxon/"Other" neutral swatches have to be re-picked alongside the hues — tuning the nine and treating the neutrals as an afterthought produced a ΔE 6.5 collision that nearly shipped.

Last reconciled moves to 2026-07-30. 275 tests still pass.

🤖 Generated with Claude Code

---

Second commit: priority menu + three corrected numbers

Reconciling the backlog against the actual KB (rather than relaying the file) found that three recorded numbers had drifted, and one large gap had never been tracked at all.

Corrections, all re-measured 2026-07-30 across 305 records:

claim in the file / issue	actual
causal-edge coverage "~65/304"	59/305 — 246 records carry no downstream edge
no growth conditions "52/295" (#183)	80/305
GTDB grounding	never tracked — 569/995 taxa (57%)

The growth-conditions gap is smaller than 80 makes it sound: 52 of those are NATURAL field communities (34 STABLE, 18 PERTURBED) that legitimately have no cultivation conditions, so the curatable slice is the 28 ENGINEERED records — and even that includes pure computational models (BioModels_…, KBase_…). The real target is roughly 24.

New: a priority menu at the top of the file ranking the six actionable items, and a separate list of blocked items with the reason each is blocked, so they stop being re-litigated at every reconcile.

Recommended next: GTDB grounding backfill (#276), filed as part of this reconcile. The schema slot and the ground-taxa-gtdb skill have both existed for a while, but coverage was never measured, so the field got filled in opportunistically during unrelated passes and is now uneven: 91 records fully grounded, 140 partially, 72 with none. The partials are the sharpest part — a record where some taxa are grounded and others aren't cannot be distinguished by a downstream consumer from one where GTDB genuinely has no equivalent. It is also the only large item on the menu that depends on nothing outside this repo: the mapping table is local, so there is no literature access, curator judgment, or sibling-repo dependency involved. And it is a correctness pass, not only coverage — the skill flags GTDB reclassifications such as NCBITaxon Agrobacterium deltae → GTDB Agrobacterium leguminum.

---

Third commit: review round

The numeric claims were re-derived independently rather than taken on trust. Every measurement held — 569/995 GTDB taxa, the 91/140/72 record split (and exactly 2 files with no taxonomy key at all, which is why it sums to 303 not 305), 59/305 causal edges, 80/305 missing growth conditions, and all 19 issue/PR cross-references. Four things needed correcting.

§1 was overstated on the blocked list. The entry claimed "every one of the 34 residuals needs a term that does not exist and that we cannot mint" — but §1 itself contradicts that twice. About 12 of the 34 are obsolete GO ids needing a curator-accepted repoint, not minting, and §1 already records a near-repoint found for sodium metasilicate. Roughly 14 are genuinely un-mintable. The gate stays blocked overall since it fails on any unresolved row, but describing the whole set as upstream-blocked buries decidable work.

The NATURAL split was ambiguous. "52 of the 80 are NATURAL" is true read by ecological_state, but a reader recomputing by community_origin gets 57 — the difference being 5 NATURAL-origin records that are ENGINEERED in state and therefore belong with the curatable set, not the field communities. Both occurrences now say "STABLE or PERTURBED" and explain the 57 explicitly.

Two smaller ones: the growth-conditions target is 25, not ~24 — only three of the 28 ENGINEERED records are pure computational models. And the JSON print is cli.py:91; line 90 is the if output_json: guard.

Three sections added for items carried too pessimistically

• Enrich records lacking members/growth conditions once full text is accessible #183 — its own "52/295" is stale (now 80/305), and the curatable slice is 25 named records. Measuring each against references_cache/: 1 has real full text, 18 have abstract stubs (0.8–5.4 KB), and 6 have nothing cached at all. Those 6 were never attempted, which is not the same as paywalled, and running cache_fulltext.py on them is the cheapest next step in the section.
• Refine best-effort ontology term remaps from the id↔label cleanup #182 — not blocked, just undecided; no tooling is missing, someone has to choose per row. Flags GO:0055114 → GO:0016491 as the one remap wrong on type grounds (biological process → molecular function) rather than merely broad, and separates this surface from §1's residuals, which are easy to conflate.
• No DOI-based full-text cache path: blocks non-PubMed OA sources (e.g. exact-pair formate data for 000068) #259 — mostly shipped in Dangling-edge detection, DOI full-text caching, and a correction to the validate-references note #260/Ingest Li 2024 exact-pair formate data into 000068; add --from-file to cache_fulltext #261. What remains is that Unpaywall is queried only to name an OA location, never to fetch it, so an OA-but-not-PMC source still costs a manual round trip. Self-contained, not blocked.

Also: #199's checkboxes were stale — the legend item and the stale-template item were both done but still unticked, so the issue disagreed with this file. Reconciled on the issue, and the surviving live community.html.j2 is now called out there so nobody deletes it by following the old instruction. #277 filed for the four dangling [[wiki-links]], which point at a machine-local memory store rather than anything in the repo.