fix(ci): repair archs4 + ELM live-data test failures

[Elarwei001]

Scoped to the ARCHS4 + ELM CI-stability fixes only. The opentargets parts of this PR have been moved to #256 per maintainer preference for one-module-per-PR review.

ARCHS4 — graceful missing-color column + deterministic order

Bug: ARCHS4 intermittently omits the color column from its tissue-expression CSV response. The previous tissue_exp_df.drop(["color"], axis=1) raises KeyError: ['color'] not found in axis when this happens, so gget archs4 -w tissue ... crashes for users on those requests.

Fix: drop(columns=["color"], errors="ignore") — drops if present, no-op if absent. The color column was never used by gget archs4 (only ARCHS4's own plotting needs it).

Secondary improvement: sort key changed from "median" to ["median", "id"] (descending median, ascending id) so equal-median tissues no longer flip order between requests. Test fixtures refreshed for the deterministic ordering.

New regression tests (network-free, mocked):

• TestArchs4MissingColor.test_tissue_missing_color_does_not_crash — main regression guard for the errors="ignore" fix
• TestArchs4MissingColor.test_tissue_with_color_still_dropped — paired companion confirming the column still gets dropped when present

ELM — retry transient setup download

gget_setup(module="elm") downloads ELM database files from elm.eu.org. Cold-cache or transient network blips occasionally fail the download and break test collection (the call happens at module-import time, before any test runs). Wrap in a 3-attempt retry with a 30s sleep between attempts so transient failures don't kill the whole opentargets/ELM suite.

Out of scope (moved or punted)

• OpenTargets synonyms HTTP 400 fix + fixture drift + expression-tests-skip: see fix(opentargets): adapt to upstream API drift + skip deprecated expressions field #256.
• baselineExpression migration (resolves gget opentargets -r expression returns empty: OpenTargets retired target.expressions (data moved to baselineExpression) #247): separate PR (or rebased version of fix(opentargets): repoint expression to baselineExpression (upstream retired target.expressions) #248).

Test plan

• python -m pytest tests/test_archs4.py -v locally — all pass (including the new TestArchs4MissingColor cases)
• python -m pytest tests/test_elm.py -v locally — passes (ELM setup either succeeds first try or after a retry)
• CI green on this PR

[codecov-commenter]

Codecov Report

❌ Patch coverage is 0% with 2 lines in your changes missing coverage. Please review.
✅ Project coverage is 7.98%. Comparing base (5cf607f) to head (d640bff).
⚠️ Report is 8 commits behind head on dev.

Files with missing lines	Patch %	Lines
gget/gget_archs4.py	0.00%	2 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##              dev    #252       +/-   ##
==========================================
- Coverage   56.14%   7.98%   -48.17%     
==========================================
  Files          29      29               
  Lines        9244    9244               
==========================================
- Hits         5190     738     -4452     
- Misses       4054    8506     +4452     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[Elarwei001]

currently opentarget change the format of its resource="drugs" response from "string" to a "DrugLabelAndSource" object. the response example looks like this:

`query: { drug(chemblId: "CHEMBL1201631") { name synonyms { label source } tradeNames { label source } } }

HTTP 200
name: INSULIN HUMAN

synonyms (first 8, now WITH source):
{'label': 'Biosulin r', 'source': 'ChEMBL'}
{'label': 'Capsulin', 'source': 'ChEMBL'}
{'label': 'Exubera (inhaled insulin human)', 'source': 'ChEMBL'}
{'label': 'High molecular weight insulin human', 'source': 'ChEMBL'}
{'label': 'Human insulin', 'source': 'ChEMBL'}
{'label': 'Humulin', 'source': 'ChEMBL'}
{'label': 'Insulin', 'source': 'ChEMBL'}
{'label': 'Insulin (human)', 'source': 'ChEMBL'}

tradeNames (first 8, now WITH source):
{'label': 'Actrapid', 'source': 'ChEMBL'}
{'label': 'Afrezza', 'source': 'ChEMBL'}
{'label': 'Exubera', 'source': 'ChEMBL'}
{'label': 'Humulin br', 'source': 'ChEMBL'}
{'label': 'Humulin r', 'source': 'ChEMBL'}
{'label': 'Inpremzia', 'source': 'ChEMBL'}
{'label': 'Ins humulin r', 'source': 'ChEMBL'}
{'label': 'Insuman', 'source': 'ChEMBL'}

-> Each entry now shows {'label': ..., 'source': ...}. So source was
always available; the earlier output omitted it only because the query
selected { label } (and gget intentionally keeps just label to
preserve its historical plain-string output).`

so, will need to change the decoding structure to capture current response labels.

test code:

#!/usr/bin/env python3
"""
Show that DrugLabelAndSource really has BOTH `label` and `source` —
we only saw `label` before because our query only selected `label`.
GraphQL returns exactly the fields you request, nothing more.

Read-only OpenTargets API call. No GitHub, no writes.
Run: /Users/elar/gget/.venv/bin/python /Users/elar/cc_output/gget/verify-opentargets-source.py
"""
import json

import requests

URL = "https://api.platform.opentargets.org/api/v4/graphql"
DRUG = "CHEMBL1201631"

# Note: now we ALSO request `source` in the sub-selection.
q = f'{{ drug(chemblId: "{DRUG}") {{ name synonyms {{ label source }} tradeNames {{ label source }} }} }}'
print("query:", q, "\n")

r = requests.post(URL, json={"query": q}, timeout=60)
print("HTTP", r.status_code)
drug = (r.json().get("data") or {}).get("drug") or {}
print("name:", drug.get("name"), "\n")

print("synonyms (first 8, now WITH source):")
for item in (drug.get("synonyms") or [])[:8]:
    print("   ", item)

print("\ntradeNames (first 8, now WITH source):")
for item in (drug.get("tradeNames") or [])[:8]:
    print("   ", item)

print(
    "\n-> Each entry now shows {'label': ..., 'source': ...}. So `source` was"
    "\n   always available; the earlier output omitted it only because the query"
    "\n   selected `{ label }` (and gget intentionally keeps just `label` to"
    "\n   preserve its historical plain-string output)."
)

[Elarwei001]

try the GENE's associate diseases, the result may change (which means opentarget's response won't be stable).

for example, if compare with the previous test case in dev, current value is update both on ID and Score:

test code:

#!/usr/bin/env python3
"""
Measure the ACTUAL drift of `gget opentargets -r diseases` for IL13
(ENSG00000169194, the gene used in test_opentargets), so we can design a
RIGOROUS test (known-ID-set + score tolerance), not a weak one.

Compares the live OpenTargets top-2 associated diseases against the OLD
fixture values that PR #250 deleted.

Read-only OpenTargets API. No GitHub, no writes.
Run: /Users/elar/gget/.venv/bin/python /Users/elar/cc_output/gget/verify-opentargets-diseases-drift.py
"""
import json

import requests

URL = "https://api.platform.opentargets.org/api/v4/graphql"
GENE = "ENSG00000169194"  # IL13

# What the OLD fixture asserted (the 171 lines PR #250 removed), top 2:
OLD = [
    {"score": 0.7297489019498119, "id": "EFO_0000274", "name": "atopic eczema"},
    {"score": 0.6642728577751653, "id": "MONDO_0004979", "name": "asthma"},
]

q = f"""
{{ target(ensemblId: "{GENE}") {{
     associatedDiseases(page: {{index: 0, size: 2}}) {{
       count
       rows {{ score disease {{ id name }} }}
     }}
}} }}
"""

r = requests.post(URL, json={"query": q}, timeout=60)
print("HTTP", r.status_code)
body = r.json()
rows = (((body.get("data") or {}).get("target") or {}).get("associatedDiseases") or {}).get("rows")
if not rows:
    print(json.dumps(body, indent=2)[:800])
    raise SystemExit("no rows — inspect raw response above")

print(f"\n{'rank':<5}{'LIVE id':<16}{'LIVE name':<22}{'LIVE score':<14}"
      f"{'OLD id':<16}{'OLD score':<14}{'Δscore':<12}{'id changed?'}")
print("-" * 110)
for i, row in enumerate(rows):
    live_id = row["disease"]["id"]
    live_name = row["disease"]["name"]
    live_score = row["score"]
    old = OLD[i] if i < len(OLD) else {"id": "-", "score": float("nan"), "name": "-"}
    dscore = live_score - old["score"]
    rel = abs(dscore) / old["score"] if old["score"] else float("nan")
    idchg = "SAME" if live_id == old["id"] else f"{old['id']} -> {live_id}"
    print(f"{i:<5}{live_id:<16}{live_name:<22}{live_score:<14.6f}"
          f"{old['id']:<16}{old['score']:<14.6f}{dscore:<+12.6f}{idchg}")
    print(f"     (relative score drift this row: {rel*100:.3f}%)")

print("\nUse this to decide: (a) the acceptable ID SET per disease (did EFO<->MONDO"
      "\nflip, or is the name stable while id changed?), and (b) a score TOLERANCE"
      "\nthat survives normal release drift but still catches wrong-data. Run on a"
      "\nfew genes / a few days to see the real drift magnitude before fixing the test.")