knowledge: 3 insights — checksum-restored identifiers need a success-path warning (1 new page), delegated mutate-and-revert leaves stale .pyc (+1 merge), batch-wide domain rule dropped as superseded

[choiyounggi]

Knowledge flush — 3 insight(s)

Claimed this run: 780cc9007d23bf67, 01cc5fdd76565581, 6db22c1e9830a187. Outcome: 1 new page, 1 amended page, 1 dropped (superseded).

Verified best-practice

1. 01cc5fdd76565581 — checksum-restored identifiers need a success-path warning → verified

• Claim: when an importer restores an identifier's lost leading zero (numeric spreadsheet cell) and accepts it because the check digit passes, it must still flag/warn on the successful repair, because a single mod-11 check digit accepts about 1/11 of arbitrary inputs.
• Sources checked:
  • https://support.microsoft.com/en-us/office/keeping-leading-zeros-and-large-numbers-1bf7b935-36e1-4985-842f-5dfa51f85fe7 — live-fetched: "Excel automatically removes leading zeros…"; formatting "will not restore leading zeros that were removed prior to formatting"; Power Query can type a column as Text at import; 15-significant-digit precision.
  • https://support.microsoft.com/en-us/excel/set-automatic-data-conversions — live-fetched: "Remove leading zeros and convert to number" automatic conversion (Excel for Microsoft 365 / 2024, Windows + Mac); default converts 00123 to 123.
  • https://pandas.pydata.org/docs/reference/api/pandas.read_csv.html — dtype for reading columns as strings.
• Reproduction: ran the real validator (src/validation/tax_code.py, weights 31/29/23/19/17/13/7/5/3, check = 10 − (sum mod 11), remainder 0 rejected) against 200,000 seeded random 0+9-digit inputs → 18,089 passed (9.04%); 200,000 random 10-digit inputs → 18,117 (9.06%). Matches the candidate's 18,242/200,000 and the analytic 1/11 ≈ 9.09%. Confirmed the originating fix exists at HEAD of the source repo (src/company_import.py: was_leading_zero_restored + "restored … verify" warning).

2. 6db22c1e9830a187 — stale .pyc after a subagent's mutate-and-revert → verified (merged into an existing verified page)

• Claim: after a mutation-then-revert verification step, clear __pycache__ before trusting the next suite run; CPython's default timestamp(+size) validation can reuse the mutant's bytecode.
• Sources: mechanism already carried and cited by backend-python-language-bytecode-cache-staleness (docs.python.org import reference — "storing the source's last-modified timestamp and size"; PEP 552; py_compile; three prior field reproductions). The candidate's case ("n/a" → "N/A") is byte-length-preserving, so it falls exactly in the page's documented same-size/same-second window — consistent, no contradiction.
• What is new: the actor who mutated was a delegated subagent sharing the worktree, so the parent session had no signal to purge. Added as one edge row + one field-reproduction source line. last_verified left unchanged (docs not re-fetched this run).

3. 780cc9007d23bf67 — evaluate per-row rejection rules against the whole batch's domain set → dropped (superseded / contradicted)

• Claim: in a batch path that fans one upstream result out to many rows, evaluate per-row rejection rules (e.g. "email domain == page domain → reject") against the batch's full set of source domains.
• Verification result: the originating codebase later reversed this. Batch extraction now writes one row per page (commit b89f8c6, "write one extracted_contacts row per page in batch extraction", issue knowledge: LLM output-cap budgeting + verification-harness reverse controls (2 of 3 insights; 1 already in flight) #13), and tests/test_ai_extractor.py::test_extract_batch_email_domain_matching_no_batch_page_stays_valid pins that batch-wide domain widening wrongly flags a row whose email matches a sibling page's domain — the exact directive this candidate proposes. The evidence paths cited in the candidate (.orchestration/archive-20260910-cv0910/reviews/…) no longer exist on disk. The directive as written is not a best-practice, so it is not ingested. (The salvageable idea — attribute each fanned-out result to its own source — would need its own candidate and sourcing.)

Existing-layer check

Pages read: backend-python-language-bytecode-cache-staleness, testing-quality-mutation-harness-file-custody, backend-common-change-impact-aggregation-layer-of-a-shared-helper, databases-schema-design-column-data-types

Also read: INDEX.md, wiki/backend/index.md, wiki/testing/index.md, wiki/qa/index.md, wiki/infrastructure/index.md (grep), wiki/security/index.md (grep); keyword sweep of wiki/ for pycache|.pyc|bytecode|checksum|leading zero|check digit|mod-11|fan-out|batch.

• Candidate 2 (pyc): bytecode-cache-staleness already owns the trigger and directive (purge __pycache__ between edit iterations; "regression persists after a clean revert" Instead-of row). Same trigger, same directive → merged (edge row + source), no new page. mutation-harness-file-custody is adjacent (restore custody, not cache) and already linked through harness-reverse-controls; no change needed.
• Candidate 1 (checksum): no page covers import-time identifier repair. column-data-types only states "national ids → TEXT, leading zeros are data" (storage side). → new page backend-common-integrations-checksum-restored-identifiers; related: to databases-schema-design-column-data-types, backend-common-integrations-consumer-required-fields, security-input-validation-at-trust-boundaries; backlink added on column-data-types.
• Candidate 3 (batch fan-out): nearest page aggregation-layer-of-a-shared-helper (same helper in two aggregation layers) — different trigger; no conflict to flag because the candidate was dropped.
• Conflicts flagged: none.

Open-PR check

Listed open heads (gh pr list --search "head:knowledge/"): #189 knowledge/choiyounggi-20260910-162026, #188 …-20260908-154412, #187 …-20260906-213635, #186 …-20260906-013856, #185 …-20260906-003745, #183 …-20260904-133717, #182 …-20260903-214027, #181 …-20260903-203836, #180 …-20260903-184706, #179 …-20260903-172728.

For each: git fetch origin <head> then git diff origin/main...origin/<head> -- wiki/, scanning the changed-file list for the target pages and the added lines for pycache|.pyc|bytecode|leading zero|checksum|check digit|fan-out|per-row|batch.

• No open head touches bytecode-cache-staleness.md or column-data-types.md. Keyword hits were unrelated (GIL bytecode switching in knowledge: 6 insights — gate EVIDENCE exit-code class (unset CLAUDE_PLUGIN_ROOT), jq . rebinding in predicates, worker-reported plan contradiction, benchmark speedup attribution, sequential-dispatch test under a thread pool, lock-protection race-window test (6 new pages, 0 merges, 0 drops) #189, per-row reorder locking in knowledge: 8 insights — OKLCH dark-surface steps, benchmark-relative signals, multi-row reorder concurrency, Prisma CONCURRENTLY, graphify viz failure, newest-transcript measurement, fake-timer TTL isolation, fallback-source alert (+3 folds onto #179/#181, 1 already-merged drop) #185, batch screenshots in knowledge: 12 insights — sequential ids across parallel workers, inbound-validation task ownership, vendor benchmark claims, element crop screenshots, WebFetch summary vs raw page, Steps-prose guarantees, split fact-check verdicts (+4 folds, 1 dup) #182, queued-row retirement in knowledge: 8 insights — SwiftPM executable-target tests, off-screen macOS capture without Screen Recording, Testcontainers reaper on Docker Desktop, ciphertext orphaned by a regenerated key, model-coupled guidance aging detector, CORS preflight probe silence, queued-candidate store check, literal-constant re-assertion (+3 folds onto #181, 1 pending-duplicate drop) #186).
• knowledge: 12 insights — workflow file that registers no triggers, workflow-authored PRs (create-PR setting, personal-repo ruleset, GITHUB_TOKEN checks, --auto), Firecrawl main-content vs contact details, testcontainers community namespace, pointer-attracted particle collapse, producer wire format in mocks (+5 merges incl. Homebrew clang SDKROOT contradiction reconciled) #187 adds backend/common/integrations/contact-details-from-scraped-pages.md (same category as the new page, different trigger: main-content scrapers dropping contact regions); knowledge: 8 insights — OKLCH dark-surface steps, benchmark-relative signals, multi-row reorder concurrency, Prisma CONCURRENTLY, graphify viz failure, newest-transcript measurement, fake-timer TTL isolation, fallback-source alert (+3 folds onto #179/#181, 1 already-merged drop) #185/knowledge: 12 insights — workflow file that registers no triggers, workflow-authored PRs (create-PR setting, personal-repo ruleset, GITHUB_TOKEN checks, --auto), Firecrawl main-content vs contact details, testcontainers community namespace, pointer-attracted particle collapse, producer wire format in mocks (+5 merges incl. Homebrew clang SDKROOT contradiction reconciled) #187 add integrations/ rows to backend/index.md — row-level index edits, no content overlap.
• Verdicts: 01cc5fdd76565581 → new; 6db22c1e9830a187 → new (merge into an existing main-branch page, no open PR carries it); 780cc9007d23bf67 → not ingested (dropped on verification, no open PR carries it either).

Routing decision

Candidate	Domain / category	Page	Action
01cc5fdd76565581	backend / common/integrations	wiki/backend/common/integrations/checksum-restored-identifiers.md	new page + backend/index.md integrations row + backlink on databases/schema-design/column-data-types
6db22c1e9830a187	backend / python/language	wiki/backend/python/language/bytecode-cache-staleness.md	+1 edge row, +1 field-reproduction source
780cc9007d23bf67	—	—	dropped: superseded by later code and regression test

No new category. integrations was chosen over security/input because the concern is repair and provenance of values from an external file format, not trust-boundary rejection. It was chosen over databases/schema-design because the loss happens before storage.