@@ -25,6 +25,115 @@ fixtures with `make fixtures-nextflow`.
2525 per step (discover or author); the partial scope is stated explicitly rather
2626 than silently dropping the remaining pipeline.
2727
28+ ## Case: nf-core/eager vs. a human-authored aDNA workflow (compare and contrast)
29+
30+ A ** reference case** , not a matching test. galaxyproject/iwc #1234
31+ nf-core/eager style ancient DNA (aDNA) analysis workflow", human-authored by Mert
32+ Aydın) targets the same source pipeline this journey converts, so it gives us a
33+ hand-built answer to hold the machine's answer against. The journey will never
34+ reproduce it, and shouldn't try — the value is in * where* and * why* the two
35+ diverge. Nothing here is a pass/fail gate; the pipeline oracle in ` eval.md ` still
36+ supplies the actual gates.
37+
38+ Three caveats to keep the comparison honest:
39+
40+ - ** Version boundary.** The PR's README cites ` nf-co.re/eager/2.5.3/ ` — eager
41+ 2.x, which is DSL1 and therefore out of scope for [[ summarize-nextflow]] . The
42+ fixture pins the DSL2 rewrite on ` dev ` (3.x). The aDNA tool vocabulary is
43+ substantially the same across the two, so the contrast is meaningful, but this
44+ is not a controlled A/B and per-step differences may be version artifacts
45+ rather than authoring differences.
46+ - ** Not in the corpus.** The PR is unmerged, so ` paleogenomics/adna-analysis `
47+ does not exist at the pinned IWC SHA. [[ compare-against-iwc-exemplar]] cannot
48+ reach it and will settle on some other exemplar; that phase's output is not the
49+ comparison. Supply the PR out-of-band by URL when doing the write-up, per the
50+ URL-not-mirror principle — do not mirror the ` .ga ` into the repo.
51+ - ** Human ≠ correct.** The PR is open and unreviewed. Where the journey diverges,
52+ either side may be the better answer.
53+
54+ - fixture: ` workflow-fixtures/pipelines/nf-core__eager ` (DSL2 ` dev ` pin; large —
55+ ~ 110 ` .nf ` files, ~ 45 nf-core modules, ~ 23 local subworkflows).
56+ - reference: < https://github.com/galaxyproject/iwc/pull/1234 > , workflow
57+ ` paleogenomics/adna-analysis ` .
58+ - expect (gated by ` eval.md ` ): eager is a toolkit, not a linear pipeline — its
59+ schema exposes ~ 15 independent ` run_* ` /` skip_* ` branch groups (preprocessing,
60+ mapping, bam_filtering, deduplication, damage_manipulation,
61+ adna_damage_analysis, genotyping, metagenomics, contamination_estimation,
62+ human_sex_determination, consensus_sequence, host_removal, …). The journey must
63+ reach a ** stated scope decision** about which branches it carries, and say so
64+ explicitly rather than silently flattening the toolkit into whichever path the
65+ template happened to take.
66+
67+ A scope decision that freezes * every* branch into one hardcoded path satisfies
68+ the letter of that and misses the point: eager's optionality is much of what
69+ eager is, and a conversion that discards all of it has answered an easier
70+ question than the one asked. So the bar is that some meaningful part of the
71+ source's branch surface reaches the user as ** run-time choice** — a workflow an
72+ analyst can steer without editing it. Which choices, how many, and by what
73+ mechanism are the journey's call, and a defensible small set beats a large one.
74+ The point is that the collapse to a single frozen path should have to be argued
75+ for, not fallen into.
76+
77+ Deliberately not prescribed: the reference keeps four user-facing controls and
78+ spends 12 ` when ` steps doing it, but that is ** calibration, not a quota** —
79+ evidence that this much optionality is reachable in gxformat2 at all. A journey
80+ that keeps three different controls, or reaches the same coverage through
81+ optional inputs rather than ` when ` gating, passes. A journey that pads its count
82+ with toggles eager's schema doesn't motivate fails on intent-fidelity — the
83+ optionality has to be earned by the source, not manufactured to hit a number.
84+ - contrast (advisory; record, don't gate):
85+ - ** Which branches survive as branches.** The PR is conditional-rich by IWC
86+ standards: 52 steps, of which 12 carry a ` when ` . It keeps four user-facing
87+ branch controls — ` Choose Read Type ` (SE/PE), ` Choose Mapper ` (BWA/Bowtie2),
88+ and two optional-input conditionals (HapMap present → ANGSD X-contamination;
89+ BED present → Sex.DetERRmine variant) — while hardcoding one pick for
90+ preprocessing (AdapterRemoval), dedup (Picard MarkDuplicates), genotyping
91+ (FreeBayes), profiling (Kraken2), and damage (mapDamage). Note the line drawn
92+ is * not* "data-flow shape stays, tool-swap goes": the mapper switch is a tool
93+ swap and survives. The kept set reads closer to "choices an aDNA analyst
94+ actually makes at runtime". Does the journey's scope decision land on a
95+ comparable set, and on what principle?
96+ - ** The conditional plumbing tax.** Galaxy has no native if/else, so eager's
97+ toggles are paid for structurally: 7 ` map_param_value ` steps turn selections
98+ and optional-input presence into ` when ` booleans, and 6 ` pick_value ` steps
99+ rejoin the branches — 13 of 41 tool steps (~ ⅓) are param plumbing, not
100+ science. The pattern is duplicate-per-branch + gate + merge (`AdapterRemoval
101+ (Single-End)` / ` (Paired-End)` , ` Map with BWA` / ` Bowtie2`). Record what the
102+ journey does here rather than scoring it against this idiom — duplicate-gate-
103+ merge is one known-workable answer, not the sanctioned one, and a cheaper
104+ route to the same run-time coverage would be a finding worth having. The tax
105+ is also the honest argument * for* a narrower scope: ~ ⅓ of the step budget
106+ going to plumbing is a real cost, and a journey that cites it while keeping
107+ fewer controls has reasoned, not given up. A worked human example of that
108+ trade priced out in full is the main reason this case earns its keep.
109+ - ** Where the human answer looks arguable.** Steps 39/40 (Sex.DetERRmine
110+ with/without BED) are * both* registered workflow outputs, as are both ANGSD
111+ paths — so any given run yields outputs that are necessarily empty. If the
112+ journey avoids that, it is not automatically wrong for differing from the
113+ reference.
114+ - ** Which branches were dropped entirely.** No circularmapper, malt/megan,
115+ metaphlan, krakenuniq, pmdtools, multivcfanalyzer/consensus, pileupcaller, or
116+ mapad in the PR. Does the journey's scope decision land on a comparable
117+ subset, and does it justify the cut?
118+ - ** Tool resolution.** Per ` eval.md ` , every tool is discovered or authored. The
119+ human found Shed wrappers for the whole aDNA set (EndorSpy, Sex.DetERRmine,
120+ MtNucRatioCalculator, mapDamage included). Where the journey chose to * author*
121+ a wrapper for something that already exists in the Shed, that's a
122+ discover-shed-tool recall gap worth filing.
123+ - ** Interface shape.** The human surfaced ~ 10 named inputs with prose-legible
124+ labels and pushed eager's flag sprawl into defaults. Compare against what the
125+ interface brief produced — count, naming, and how much of the ~ 200-param
126+ surface leaked through.
127+ - ** Test strategy.** The PR ships a Planemo suite on sub-sampled ` JK2067 ` /
128+ ` hs37d5 ` / HapMap data. Compare against what
129+ [[ nextflow-test-to-galaxy-test-plan]] → [[ implement-galaxy-workflow-test]]
130+ produce — did the journey reach real aDNA fixtures, or synthesize? This is
131+ the case most likely to expose the pipeline's known test-data gap: the spine
132+ has no [[ find-test-data]] phase, so ` implement-galaxy-workflow-test ` takes
133+ ` test-data-refs ` with nothing upstream producing it (the validator warns on
134+ exactly this). aDNA reference data is heavy and non-obvious; if the gap bites
135+ anywhere, it bites here.
136+
28137## Case: branch-control fidelity
29138
30139- fixture: a Nextflow summary with a skip-style branch control (e.g.
0 commit comments