knowledge: 8 insights — Python class introspection, PG derived-table SubLink duplication, browser console-capture gaps, extension reload, degenerate expectation sets, ORM-generated test schema, APFS recovery

[dch0202-rsquare]

Knowledge flush — 8 insight(s)

8 queued candidates → 7 new pages (two candidates share one mechanism and were
merged into a single page), 7 existing pages amended with reciprocal related:
links, 5 domain indexes updated, log.md appended.

Verified best-practice

1. inspect.getsource on a class from a path-loaded module → verified

Claim: a module loaded via spec_from_file_location/module_from_spec/exec_module
without sys.modules registration makes class introspection fail with
TypeError: … is a built-in class, while function/method introspection works.

How verified: reproduced locally 2026-08-18 on CPython 3.9.6, 3.11.13, 3.13.11,
3.14.6 — the headline behaviour is identical on all four: class → TypeError: <class 'aw.Handler'> is a built-in class; m.Handler.do_GET and m.select → source returned; after
sys.modules[spec.name] = m the class call returns source. Mechanism read directly from
the shipped inspect.py (3.14.6) getfile: the class branch resolves through
sys.modules.get(object.__module__), the function branch falls through to
object.__code__.co_filename. The registration line was confirmed to be part of the
standard library's own documented recipe.

Sources checked (opened this session): docs.python.org/3/library/importlib.html
("Importing a source file directly" recipe, sys.modules[module_name] = module),
docs.python.org/3/library/inspect.html, docs.python.org/3/reference/import.html,
local inspect.py source.

2. Unpacked Chromium extension reload → verified

Claim: after editing a content script, the host-page refresh alone re-injects the
cached copy; the extension itself must be reloaded too.

How verified: Chrome's official Hello-World tutorial was opened and states "After
saving the file, to see this change in the browser you also have to refresh the
extension", with its reload table listing content scripts as "Yes (plus the host page)".
chrome.runtime.getManifest() (used for the running-build marker directive) was
confirmed on the official runtime API page. The #dev-reload-button shadow-root path is
labelled in-page as a measured internal detail, not a documented API, with instructions
to re-derive it when it stops matching. Field evidence: okta-autofill, page-refresh-only →
field stayed empty; extension reload → flow completed.

Sources checked (opened): developer.chrome.com get-started tutorial,
develop/concepts/content-scripts, reference/api/runtime.

3+8. Browser-automation console capture gaps → verified (merged page)

Claim A (candidate 8): a collector attached after navigation loses load-time records.
Claim B (candidate 3): extension content-script logs live in an isolated execution
context that a main-world collector never reports.

How verified: Claim A independently reproduced this session (Aside CLI
1.26.810.1915 / daemon 1.26.818.1059) against a local python3 -m http.server page whose
inline script logs one log and one error: openTab → []; after sleep(800) → [];
after console.clear() + reload() + sleep(300) → both records. Claim B's mechanism is
documented — Chrome content-scripts page ("An isolated world is a private execution
environment that isn't accessible to the page…") and CDP Runtime.consoleAPICalled
carrying executionContextId, "Identifier of the context where the call was made"; the
default/isolated/worker distinction is carried in executionContextCreated's auxData
("Embedder-specific auxiliary data"), and the page now says so rather than presenting it
as a documented field of its own. The tool-specific miss stays labelled as a field
measurement.

Correction made: a drafted playwright.dev/docs/api/class-page citation was dropped
— the fetch did not surface the console event section, so the claim could not be
confirmed from it.

Sources checked (opened): chromedevtools.github.io Runtime domain,
developer.chrome.com content-scripts. (playwright.dev opened, not confirmable → removed.)

4. Correlated SubLink duplicated by derived-table pull-up → verified

Claim: a derived table carrying no aggregate/LIMIT/DISTINCT/set-op is flattened, its
select-list SubLink is substituted into every referencing aggregate, and PostgreSQL does
not CSE SubPlans — so N outer references cost N evaluations per row; OFFSET 0,
MATERIALIZED, or LATERAL fix it.

How verified: the harvested evidence was second-hand (a worker's measurement the
harvesting session could not re-run). Reproduced from scratch this session on a
throwaway local cluster, PostgreSQL 16.11 (Homebrew), 5 driving rows:

Shape	Plan
Unfenced derived table, read by bool_or + sum	SubPlan 1 + SubPlan 2, each loops=5 (10 evaluations)
OFFSET 0 inside the derived table	one SubPlan, loops=5
WITH … AS MATERIALIZED	one SubPlan under a CTE node
Plain WITH (not materialized)	two SubPlan nodes — a plain CTE is not a fence
LEFT JOIN LATERAL	no SubPlan — Nested Loop Left Join
Derived table with GROUP BY	one SubPlan
1 / 2 / 3 outer references	1 / 2 / 3 SubPlan nodes

The plain-CTE and reference-count rows are additions the candidate did not contain.
Doc quotes obtained verbatim: OFFSET 0 "is the same as omitting the OFFSET clause"
(queries-limit), and the folding/MATERIALIZED/push-down-restriction sentences
(queries-with). The pull-up predicate is cited by location (is_simple_subquery() in
prepjointree.c) and explicitly marked as measured-not-quoted, because the only doxygen
rendering available returned a paraphrase rather than the source comment.

Sources checked (opened): postgresql.org queries-limit.html, queries-with.html;
local EXPLAIN / EXPLAIN ANALYZE runs.

5. Expectation sets with one distinct value → verified

Claim: when every case expects the same literal, a hardcoded constant at the assembly
point is observationally identical to the wired computation, so adding assertions cannot
kill it; a delete probe only proves key presence.

How verified: reproduced with a minimal control (CPython 3.9.6, unittest):
3 cases all expecting "HAS_VACANCY" → constant mutant survived (0 failures);
adding one case expecting "UNSURVEYED" → same constant killed (1 failure); wired
baseline green in both sets; a delete probe reddened via KeyError, i.e. on key presence
alone. Field measurement (rtb-unified NEWRTB-2786, 2046-test api suite) retained as the
production instance.

Correction made: the Stryker and PIT quotes were re-fetched rather than inherited —
the real sentences are "When all tests passed while this mutant was active, the mutant
survived. You're missing a test for it." and "Survived means the mutation was not detected
by the covering test." A third drafted citation (testing.googleblog.com) was dropped as
unopened.

Sources checked (opened): stryker-mutator.io mutant-states-and-metrics, pitest.org
basic_concepts; local reproduction.

6. ORM-generated test schema hides model-vs-DB drift → field-tested

Claim: with ddl-auto: create-drop/create/update the test schema is generated
from the entity model, so an entity-vs-database drift cannot exist there and "add a
reproducing test" is unachievable; verification must move to a migration-built DB with
validate or an information_schema gate.

How verified — and why not verified: the knobs are documented (Spring Boot: JPA
databases "are automatically created only if you use an embedded database"; the
ddl-auto ↔ hibernate.hbm2ddl.auto mapping; "If you are using a higher-level database
migration tool, like Flyway or Liquibase, you should use them alone to create and
initialize the schema"; Jakarta @Column.nullable = "(Optional) Whether the database
column is nullable"). The consequence — that the drifted state is unconstructible in a
generated schema — follows from those but was not executed here (no JVM reproduction
run), so the page stays field-tested on the manage-repo observation rather than claiming
a measurement it does not have. A Hibernate User Guide fetch for the hbm2ddl.auto value
table returned a truncated section, so the value enumeration was first folded into the
Spring Boot bullet — which does not enumerate the values. The cross-check caught that;
the enumeration now cites Hibernate's org.hibernate.tool.schema.Action javadoc directly
(NONE/CREATE_DROP/UPDATE/VALIDATE with their legacy hbm2ddl.auto names), which
also supplies the "Drop the schema and then recreate it on SessionFactory startup"
semantics the page's mechanism rests on.

Sources checked (opened): docs.spring.io reference/data/sql.html,
how-to/data-initialization.html, jakarta.ee @Column javadoc.

7. Deleted-file recovery on macOS/APFS → field-tested

Claim: check TRIM and APFS snapshots before recommending any recovery tool, work copy
sources in fidelity order, and read the exact path from an app's bookmark blob when the
remembered name is wrong.

How verified: the probe commands were run locally 2026-08-18 (macOS 15, Darwin
25.5.0, APPLE SSD AP0512Z): system_profiler SPNVMeDataType → "TRIM Support: Yes";
tmutil listlocalsnapshots /System/Volumes/Data runs and prints its header with no
snapshots. Man pages quoted: tmutil(8) listlocalsnapshots/localsnapshot, trimforce(8)
("By default, TRIM commands are not sent to third-party drives" — the basis for the
third-party edge case). The strongest claim — that carving is not a viable path with TRIM
active — has no Apple statement behind it, so the page keeps field-tested and phrases
the directive as routing to copy sources rather than asserting impossibility.

Correction made: two drafted Apple URLs (a Disk Utility support page and the NSURL
bookmarkData developer page) were dropped — the developer page is JS-rendered and
returned no body, and neither was confirmable; the page now cites the local man pages and
measurements instead.

Existing-layer check

Routing started at INDEX.md, then each domain index.md; every category directory that
could plausibly own a candidate was listed and the overlapping pages were opened in full.

Pages read: testing-quality-checks-that-cannot-pass, testing-quality-generated-sql-property-assertions, testing-quality-source-text-wiring-assertions, testing-quality-tests-that-cannot-fail, testing-quality-unasserted-return-fields, testing-quality-default-values-under-test, qa-environments-test-environment-parity, backend-java-jpa-entity-mapping, platforms-tools-version-keyed-artifact-cache, backend-python-language-bytecode-cache-staleness

Overlaps found and how they were resolved

Candidate	Nearest existing page	Verdict
1 (inspect/class)	backend-python-language-bytecode-cache-staleness (same "edited/loaded source vs what runs" family), testing-quality-checks-that-cannot-pass (an always-red check)	New page — neither carries the resolution-route mechanism; cross-linked both ways to the first and inline to the second
2 (extension reload)	platforms-tools-version-keyed-artifact-cache (cache serves the old artifact)	New page — that page is version-string cache invalidation for a distribution system; this one is a per-file-type reload matrix. Reciprocal link added
3+8 (console capture)	qa-environments-headless-browser-bot-blocking (only other browser-environment page); platforms-processes-tool-diagnostics-without-a-failing-exit-code (empty result ≠ clean)	One new page for both candidates — same trigger ("a console reading is about to become a verdict"), two disjoint blind spots. Reciprocal link added to the bot-blocking page
4 (SubLink duplication)	testing-quality-generated-sql-property-assertions — its step 4 said an aggregate-occurrence count "doubles as the single-evaluation regression guard for a correlated subquery"	New page + a content amendment to the old one. As written the two pages did conflict for a reader who loaded only the old one: a string assertion cannot see planner-level duplication. That step is now scoped to a textual second evaluation and routes planner-level duplication to the new page. (This row previously read "no conflict — the two agree"; the cross-check below refuted that, and the amendment is the fix)
5 (degenerate expectations)	testing-quality-default-values-under-test step 1 covers the same degeneracy scoped to a constructor/factory default; unasserted-return-fields covers fields no assertion reads	New page for the general expectation-set form, with explicit edge-case rows routing the default-parameter form and the unread-field form to those two pages. Reciprocal links added to both
6 (ORM test schema)	qa-environments-test-environment-parity (parity inventory), backend-java-jpa-entity-mapping	New page — parity page owns the release-decision inventory, this owns the "what can this test level reproduce" decision. Reciprocal link added to the parity page
7 (APFS recovery)	none in platforms/filesystems	New page

Conflicts flagged: none. No candidate contradicted an existing directive.

New categories: none — all seven pages landed in existing categories.

Open-PR check

gh pr list --repo choiyounggi/dev-loop --state open --limit 50 returned zero rows
(also with --search "head:knowledge/"). There are no knowledge/* heads in flight, so
no candidate could overlap a pending PR.

Candidate	Overlapping open head	Verdict
1 inspect/class	none	new
2 extension reload	none	new
3 isolated-world console	none	folded into candidate 8's page (same-flush merge, not an open-PR fold)
4 SubLink duplication	none	new
5 degenerate expectations	none	new
6 ORM test schema	none	new
7 APFS recovery	none	new
8 console buffer timing	none	new

Routing decision

#	Target	Confidence
1	backend/python/language/source-introspection-of-a-dynamically-loaded-module.md	verified
4	databases/query-optimization/repeated-sublinks-in-a-pulled-up-derived-table.md	verified
3+8	qa/environments/browser-console-capture-gaps.md	verified
2	platforms/tools/unpacked-extension-source-reload.md	verified
5	testing/quality/expectation-sets-with-one-distinct-value.md	verified
6	testing/strategy/orm-generated-test-schema.md	field-tested
7	platforms/filesystems/deleted-file-recovery-on-apfs.md	field-tested

Routing notes:

• 2 → platforms/tools rather than frontend: the lesson is a developer-loop/tooling
  fact (which artifact the runtime is serving), not web-UI code; frontend/ has no
  extensions category and creating one for a tooling lesson would split the
  cache-staleness family across domains.
• 3+8 → qa/environments rather than testing: the subject is judging a running
  system's output as a release/QA verdict, not writing automated test code.
• 6 → testing/strategy rather than qa/environments: the decision it drives is "which
  level can hold this defect", which is test-level-choice's neighbourhood; the parity
  page keeps the release-decision framing and now links here.

Verification of this change

• node scripts/wiki-structure-checks.js wiki → pages: 249, indexes: 13, findings: 0
• node scripts/wiki-lint-prohibitions.js wiki → directives: 71, compliant: 71,
  violations: 0 (1 pre-existing info row in infrastructure/config, untouched here)
• Gate-reads-my-files control: a bare prohibition inserted into the new console page
  moved the linter to violations: 1; restoring from a cp backup returned it to
  violations: 0 with the mutation marker absent and the file byte-identical (cmp).
  Without this the green run would not have been evidence the new pages were scanned.
• Body lengths 63–72 lines, all within the ≤120 rule.
• Not run locally: the bats tests/ suite (bats is not installed on this machine) —
  CI runs it on this PR.

Cross-Check

An independent adversarial reviewer (separate claude -p process, --permission-mode plan,
no shared context) was run against this branch before the PR, tasked only with source
fidelity, overclaim, cross-page contradiction, report accuracy and AGENTS.md compliance.
Verdict: REVISE — 4 major, 8 minor. All were addressed, and each major was re-checked
by me independently rather than taken on the reviewer's word:

#	Finding	Independently re-checked	Fix applied
1	The PostgreSQL source path was wrong — optimizer/plan/prepjointree.c does not exist	gh api repos/postgres/postgres/contents/…: the file is under optimizer/prep/, not plan/	Path corrected. The reviewer also noted the predicate rejects sortClause; I measured it (derived table with ORDER BY → one SubPlan) and added both the decision row and the measurement arm
2	The __main__ edge case is version-dependent, while the page claimed four versions behaved "identical"	Re-measured with python -c (so __main__ has no __file__): 3.9.6 → TypeError, 3.11.13 / 3.13.11 / 3.14.6 → OSError: source code not available	Row scoped by version with the measurement; the "identical on all four" claim narrowed to the headline behaviour
3	The ddl-auto value enumeration was attributed to a Spring Boot page that does not enumerate it	Re-read my own fetch: the enumeration came from the fetch tool's summary, not from the page itself	Split out and cited to Hibernate's Action javadoc, opened this session
4	Cross-page contradiction with generated-sql-property-assertions, which this report had declared conflict-free	Read both side by side — the old step 4 does overreach for a reader loading it alone	Old step 4 scoped to textual duplication and cross-linked; the Existing-layer row above corrected

Minor findings applied: version-scoped the planner claim to the measured 16.11; labelled
the push-down rationale as inferred rather than cited; "three assertions" → "three cases"
in the expectation page (the very axis that page teaches); rewrote three source bullets
that had dropped their source's hedge or scope (getfile vs getsource argument list,
getManifest() wording, executionContextCreated auxData); dropped two weakly-adjacent
related: ids from the APFS page and replaced its one hedged cell with the stated
condition; corrected the lint numbers and the body-line range in this report.

Not changed, with reasons: the "Also when" trigger shape flagged on the console page is
house convention (71 of 252 pages use it) and both halves share one decision point;
.dev-loop/CROSSCHECK_FINDINGS.md and .dev-loop/fold-note-73.md are pre-existing
untracked scratch from an earlier flush — this commit stages explicit paths only, so
neither ships, and neither is mine to delete.

The reviewer stated it could not verify the off-checkout evidence (the Aside runs, the
okta-autofill observation, NEWRTB-2786, the manage repo, the PG 16.11 EXPLAIN runs, and
the lint mutation control) and neither confirmed nor refuted it — those rest on the
measurements recorded in the pages.

[dch0202-rsquare]

Cross-Check: independent adversarial claude -p review (separate process, --permission-mode plan, no shared context) was run on this branch before the PR — verdict REVISE, 4 major + 8 minor. All applied; each major was re-verified independently (wrong PostgreSQL source path confirmed via gh api, the __main__ branch re-measured across 3.9/3.11/3.13/3.14, the ddl-auto enumeration re-sourced to Hibernate's Action javadoc, and the cross-page contradiction with generated-sql-property-assertions fixed by amending that page's step 4). Details in the report's Cross-Check section (pushed as a follow-up commit, since I cannot edit the PR body from the fork account).

Note: the dev-loop:knowledge label could not be applied — the fork account has no permission to update PRs on this repo.