knowledge: import-time test coupling, empty-data surveys, call-site census, shell off-switches

[choiyounggi]

Opened via the GitHub MCP tool because the local gh CLI token is expired (HTTP 401 on api.github.com/graphql). Commit is authored by dch0202 <dch0202@infobank.net>; branch pushed over SSH. Please add the dev-loop:knowledge label if it does not appear — the label could not be set from here.

Knowledge flush — 4 insight(s)

Drained from ~/.dev-loop/queue/ (4 pending rows across 3 sessions / 3 repos:
chungyak-alimi, track-b-ranking, linkly). Four pages created, one new
category, twelve existing pages back-linked. No page overwritten, no conflict found.

Verified best-practice

1. Import-time I/O makes a "pure function" unit test infrastructure-bound

Claim — a pure function reached by importing a module whose top level runs
init_db()/connect is not DB-free; a module-level pytestmark = pytest.mark.skipif(...)
cannot prevent it, because collection imports the module body top-to-bottom and the
app import above that line has already run.

Sources checked

• https://docs.pytest.org/en/stable/how-to/skipping.html — confirms the three real
  escape hatches: pytest.importorskip at module level, pytest.skip(reason, allow_module_level=True), and pytestmark = pytest.mark.skipif(...) for a module.
  (Read from the upstream doc source doc/en/how-to/skipping.rst on
  pytest-dev/pytest@main after docs.pytest.org returned HTTP 429.)
• https://docs.pytest.org/en/stable/example/pythoncollection.html — collect_ignore /
  collect_ignore_glob; states pytest imports files matching the discovery patterns,
  which breaks on files that raise on import. This is the mechanism the insight names.
• https://docs.python.org/3/reference/import.html — module body executes on first import.

How verified — official docs plus a local reproduction (2026-08-05): a module
whose first line prints a side effect still prints it when only its pure function is
imported (from modside import pure_fn → side effect printed, then pure_fn(2) == 4).

Refinement made during verification — the queued directive said skipif markers are
"무력" (powerless). That holds for the ordering reason only; the docs show that an
allow_module_level skip placed above the import, importorskip, and
collect_ignore all do work. The page states the ordering rule ("place the guard above
the import it protects") rather than the blanket claim.

Confidence: verified

2. Confirm volume before reading a distribution query

Claim — GROUP BY/DISTINCT over an empty table return zero rows without error,
which reads as "no values to normalize"; count(*) returns a row containing 0 and is
the only unambiguous shape. When the count is 0, derive the rule from the writers and
fixtures instead, and record the substitution in the artifact.

Sources checked

• https://www.postgresql.org/docs/current/functions-aggregate.html — "except for
  count, these functions return a null value when no rows are selected. In
  particular, sum of no rows returns null, not zero as one might expect, and
  array_agg returns null rather than an empty array".
• https://greatexpectations.io/blog/exploring-data-quality-volume/ — volume (row count)
  as a first-class data-quality dimension; undetected volume anomalies "skew analyses
  and lead to flawed decision-making". Supports "check volume first" as established
  practice rather than a local habit.
• https://www.postgresql.org/docs/current/tutorial-agg.html was checked first and does
  not state the empty-input behaviour; it is therefore not cited.

How verified — reproduced 2026-08-05 in sqlite3 :memory: over an empty table:
SELECT area_nm, count(*) … GROUP BY area_nm → 0 rows; SELECT count(*) → one row
0; SELECT max(area_nm) → one row NULL. Matches the PostgreSQL-documented
semantics, so the page is written engine-neutral.

Confidence: verified

3. Enumerate call sites by call target, not by parameter name

Claim — grepping param= cannot find callers that pass the argument positionally,
so a keyword-based census silently under-reports; test helpers compound this by
reproducing the old shape at many sites while appearing once in a call-target search.

Sources checked

• https://docs.python.org/3/reference/expressions.html#calls — "If there are N
  positional arguments, they are placed in the first N slots"; keyword arguments are
  matched by identifier. This is the mechanism: a positional call site contains no
  parameter name, so no keyword pattern can match it.
• https://libcst.readthedocs.io/en/latest/codemods.html — a codemod is "an automated
  refactor that can be applied to a codebase of arbitrary size"; CST transforms match
  call nodes rather than text. (Definition confirmed via the upstream
  docs/source/codemods.rst; readthedocs returned HTTP 429.) The page cites LibCST
  only for the tool-choice row. A widely-repeated "regex is insufficient for call-site
  refactors" line traces to a Medium post I could not fetch, so that wording is not
  asserted anywhere in the page.
• The originating session's own failure record: a repo_rows= census reported 13 hits,
  the real run was Ran 472 tests / FAILED (failures=11), all in test_backend.py,
  plus a rows_for() helper feeding 5 further sites.

How verified — language-reference semantics (positional binding by position) plus
the reproduced failure above. The mechanism generalizes to any language with positional
calls, so applies_to: [general].

Confidence: verified

4. ${VAR:-default} discards an empty override

Claim — passing VAR= to turn a feature off is silently ignored when the script
reads ${VAR:-default}, because the colon form substitutes the default for null and
unset. Pass a sentinel the script's own validation rejects, or change the read to
${VAR-default}.

Sources checked

• https://pubs.opengroup.org/onlinepubs/9699919799/utilities/V3_chap02.html (POSIX
  2.6.2) — "use of the <colon> in the format shall result in a test for a parameter
  that is unset or null; omission of the <colon> shall result in a test for a parameter
  that is only unset", with the full four-operator behaviour table.
• https://www.gnu.org/software/bash/manual/html_node/Shell-Parameter-Expansion.html —
  "Omitting the colon results in a test only for a parameter that is unset."

How verified — reproduced 2026-08-05 in bash: with v="", ${v:-tmux} → tmux
while ${v-tmux} → empty; with v unset both → tmux.

Confidence: verified

Existing-layer check

Pages read in full before writing — AGENTS.md, INDEX.md, the testing,
databases, platforms, debugging, qa and backend/python domain indexes,
testing/strategy/test-level-choice, testing/data/test-data-and-isolation,
databases/query-optimization/existence-and-count-checks,
databases/schema-design/requirements-to-tables, qa/process/regression-scope,
platforms/shells/portable-shell-scripts, plus the full 135-page inventory.

Overlaps found, and what was done

Existing page	Overlap	Decision
testing/strategy/test-level-choice	Its edge case "logic buried in a controller… extract it" shares the extraction remedy	New page. Different trigger (import-time I/O, failing at collection) and a different decision set (importorskip / module-level skip / collect_ignore). Cross-linked both ways
testing/data/test-data-and-isolation	Owns state leakage between tests	Not merged. Import-time coupling is a dependency-surface question, not fixture ownership; the new page hands off to it for the module-singleton case
databases/query-optimization/existence-and-count-checks	Both discuss count(*)	Not merged. That page answers "how do I write an existence check cheaply"; the new page answers "what does an empty result prove". No conflict — the count(*)-first rule is about evidence, not cost. Cross-linked
databases/schema-design/requirements-to-tables	Deriving structure from requirements	Not merged. Requirements-driven vs data-driven derivation. Cross-linked
qa/process/regression-scope	Its Integration ring: "test every consumer of that contract, not a sample"	Complementary. regression-scope says which consumers must be tested; the new page says how to find all of them. Strongest link in the batch; cross-linked both ways
platforms/shells/portable-shell-scripts	Mentions "${OPT:-}" once, as a set -u workaround	Not merged. That page's trigger is cross-machine portability; nothing in it covers colon-vs-no-colon semantics or the caller-side override problem. The new page's set -u edge case defers to it. Cross-linked
platforms/environment/unicode-text-matching	Also a "grep returns fewer hits than reality" page	Distinct and deliberately not linked — encoding/normalization cause vs call-syntax cause, different domain; a link would create a misleading route

Conflicts flagged: none. No existing directive is contradicted by any of the four.

Related-links added (both ways): 12 existing pages gained the new ids —
testing/strategy/test-level-choice, testing/data/test-data-and-isolation,
backend/python/language/mutable-state-traps,
databases/query-optimization/existence-and-count-checks,
databases/schema-design/requirements-to-tables,
databases/schema-design/nullability-and-defaults, qa/process/regression-scope,
testing/quality/checks-that-cannot-pass, debugging/methodology/verify-the-fix,
platforms/shells/portable-shell-scripts, platforms/environment/path-resolution,
platforms/processes/non-interactive-cli-invocation.

Routing decision

#	Target page	New category?
1	testing / strategy / import-time-side-effects	No
2	databases / data-survey / surveying-live-data-for-a-rule	Yes — data-survey
3	qa / process / enumerating-call-sites-of-a-changed-signature	No
4	platforms / shells / disabling-a-feature-through-an-environment-variable	No

1 — testing/strategy. The decision this page forces is structural: what level the
test really sits at, and whether the function moves or the whole module gets gated.
strategy owns level/structure decisions; data owns fixtures and state, which is not
the question.

2 — databases/data-survey (new category). Existing categories are indexing,
query-optimization, schema-design, operations, sqlite, transactions. None
covers reading live data as evidence for a decision: query-optimization is about
cost, schema-design derives structure from requirements rather than from rows. Each
was re-checked under its alternative reading before creating the category, per
wiki-ingest step 3. databases is the right domain because the artifact queried and
misread is the database. Root INDEX.md and the domain index were widened to route in.

3 — qa/process. The one judgement call worth your attention. The queue's own
hint was domain: testing and the observed failure was 11 broken tests — but the
page's subject is enumerating what a change touches, which is
qa/process/regression-scope's question one level lower (a code-level census feeding
its Integration ring). Alternatives considered and rejected: testing/* (the practice
governs production call sites too, not test authoring); testing/quality (sibling to
the grep-gate pages, but those are about checks, not censuses); debugging (nothing
is being diagnosed); backend/common (the practice is language- and tier-agnostic).
If you prefer it under testing, this is the page to move — say so on review and it
moves with its links.

4 — platforms/shells. Category already exists and the case is squarely shell
semantics. Folding it into portable-shell-scripts as an edge-case row was considered
and rejected: that page's trigger is "runs on machine A, fails on machine B", which
does not match "my override was ignored", and one-case-per-page applies.

Verification of the wiki edit itself

• Body lengths: 56 / 65 / 58 / 58 lines — all under the 120-line cap.
• All related: ids across all 135 pages resolve; all inline [page-id] refs resolve.
• All INDEX.md and domain-index links resolve to existing files.
• No banned vague qualifiers in the new pages.
• log.md carries the ## [2026-08-05] ingest | … entry.