feat(research-areas): refresh all ResearchAreas pages for matrix completeness

[benjibromberg]

What & why

The ResearchAreas/*.md deep-dives behind the Papers.md matrix had drifted into two tiers: four mature pages and four near-empty stubs that covered only a fraction of their matrix columns (e.g. Cellular Engineering's column spans ~32 papers; its prose covered ~6). This PR brings all eight pages up to the mature template and reconciles each against its matrix column, so every paper assigned to a column is now represented in that column's page.

How it was built (grounding + adversarial review)

Every factual claim is grounded in the paper's full text, pulled from the locally-synced caail Zotero group library (localhost:23119) rather than abstracts — per the repo's AI-agent convention. Every drafted/edited page then passed the repo's two read-only adversarial reviewer subagents before commit (writer ≠ reviewer):

• caail-citation-reviewer — every (… ref #N) anchor resolves to the right Papers.md ID with matching author/year, and every Software.md/Databases.md/Datasets//Talks.md cross-link resolves.
• caail-claim-reviewer — every count, metric, method description, and identity claim checked against source full text.

The review caught and fixed real issues, e.g. a contradicted CFD cross-validation figure (Bioprocess #29), a GA-objective drift (#31), an unsupported "first" superlative (Media #1), incomplete enumerations (Sensory #171, AIEval #164 SciHorizon-Gene), and a King-2004 headword ("Adam" → "The Robot Scientist").

Per-page coverage

Page	Work	Matrix cells now covered
Scaffolding	full rewrite to template	refs #19, #20, #34, #35
Media Optimization	full rewrite	#1–3, #15–18, #21, #23–25, #58, #169, #170
Bioprocess Control	full rewrite	#7, #29–33, #59, #61, #62
Cellular Engineering	full rewrite (~32 papers)	single-cell ML, biological foundation models, perturbation prediction, genetic design, agents
Sensory Prediction	+taste/off-flavor + texture	added #102–105, #171, #195
AI Evaluation & Benchmarking	+frontier & gene benchmarks	added #155–159, #164, #196
AI Tooling / Methodology	+13 papers across clusters	added #133, #151–154, #160–163, #166, #167, #182, #197
Metabolic Modeling	verified (no matrix column)	all cross-links confirmed; no changes needed

Verification

• Anchor integrity: all Papers.md#N anchors across the 8 pages resolve to existing IDs (0 dangling).
• Column completeness: every matrix-column paper now appears in its page's prose (0 missing across all 7 column-backed pages).
• Site build: pnpm --dir site build succeeds — all 8 /research-areas/* pages render.

Follow-ups (out of scope here — not modifying Papers.md)

Grounding surfaced a few likely matrix-placement nuances to consider in a separate PR: #34 Andrews 2025 sits in the SVM row but uses a genetic algorithm + DL surrogate (no SVM); #6 DNABERT sits in Deep Learning though it is a masked-LM model; #60 Mathieu and #145 MetaGEM are placed among foundation models but are an interactome perspective and a metabolic-model reconstruction respectively. The prose describes each accurately; the matrix cells were left untouched.

🤖 Generated with Claude Code

[benjibromberg]

Updated: reconciled against main (matrix audit #33 + field-gap #32)

Since this PR was opened, main advanced ~40 commits — most consequentially the #33 matrix-classification audit and the #32 field-gap additions, which rewrote Papers.md. Merged main in (clean — zero file overlap) and reconciled all eight pages against the new matrix.

Correctness

• Removed 4 dangling anchors whose refs the audit deleted: #22 (Lao), #151/#152 (Gu CHI/HCI), #163 (MoleCode).
• Renamed the Bioprocess page → Bioprocess & Scale-Up (H1, lede, and all cross-links) to track the renamed column.
• Reframed reclassified refs: #60 (now a cultivated-meat perspective, not "the one column study"); #90/#126/#62 corrected from Cell-Eng/Bioprocess cells to their AI Tooling cells; #68 recharacterized (LLM literature-extraction + hybrid GEM/DL predictor, not a RAG design workflow).

Completeness — ~39 newly-added column papers integrated (each Zotero-full-text-grounded and gated by caail-claim-reviewer, writer ≠ reviewer):

Page	New refs
Media Optimization	#210, #211, #212, #213
Cellular Engineering	#235, #218, #63, #66, #68
Bioprocess & Scale-Up	#200–#203 (new RL cluster), #204, #205, #206–#208, #209, #27
Scaffolding	#214, #215, #216, #217
Sensory Prediction	#196, #236
AI Tooling	#4, #198, #199, #219, #220, #221, #222, #223, #224, #126, #90
AI Evaluation	#225, #53, #57

Adversarial review caught and fixed real errors before commit, e.g. #208 ("Raman > capacitance" is only true for viability — capacitance wins VCD), #209 ("coefficients above 0.99" → held-out test 0.99; validation was 0.9856), and the #68 RAG mischaracterization.

Verified: 0 dangling anchors · every matrix-column paper represented in its page · pnpm --dir site build succeeds (38 pages).

[benjibromberg]

Follow-up: MeatScan recatalogued as a dataset, not a benchmark (3bb81d6)

Per review feedback — MeatScan (#196) is a Data in Brief image dataset, not an AI/ML benchmark, and it lives in the Sensory Prediction matrix column (CNN), not AI Evaluation.

• Removed from Datasets/Benchmarks.md (the benchmark-datasets page); canonical home stays Datasets/Cow.md.
• Removed the stray bullet from the AI Evaluation deep-dive's "Domain-specific predictive benchmarks" section.
• Fixed a resulting double-count (it was tallied both as a Cow.md inventory row and a Benchmarks.md H2): catalogued-dataset total 147 → 146, benchmarkEntries 15 → 14, with the three pinned parser test ground-truths updated.

Unchanged: its Sensory Prediction mention (already framed as a dataset → Datasets/Cow.md) and its Papers.md #196 entry.

Note: this commit deliberately touches a canonical Datasets/ file (the no-canonical-files boundary was waived for this specific correction). Verified: pnpm --dir site test (296 pass) and pnpm --dir site build (38 pages) both green.