Wave 19: structural integrity — shift clamps, cache/equals fidelity, formula-record + bytes preservation, lint teeth by arcaputo3 · Pull Request #436 · TJC-LP/xl

arcaputo3 · 2026-07-22T22:01:43Z

Burn-down wave 1 of 3 (plan of record 2026-07-22): the corruption/degradation class from the field-gotcha audit, run as five worktree-isolated TDD clusters with pipelined adversarial review (one rework round, structural cluster — 8 findings, all resolved and re-approved).

What's in it

structural (structural: insert/delete rows/cols rewrite every <f> with a leading '=' and discard every cached <v> #427, structural: range shifts have no upper clamp — full-height ranges overflow past row 1048576 / col XFD (Excel refuses the file) #428, structural: Excel-authored dataValidations, print areas, and tables/autoFilter never shift on insert/delete #429 — serial): insert/delete-rows/cols no longer poison files. Formulas re-print equals-free and keep their cached <v> through shifts (openpyxl reads =A4*2, not ==A4*2; sheets no longer blank until recalc). Range shifts clamp at the sheet edge (Sheet.shiftRangeOnAxis, FormulaShifter) — the A144:XFD1048578 Excel-refuses-file class is closed. Excel-authored dataValidations (showInputMessage/showErrorMessage-stamped), pageSetup print areas, tables, and preserved autoFilter now shift with the edit via the new design-panel-directed SqrefShift engine; the writer's clean-passthrough correctly goes dirty.
formula-records (ooxml: t="array" and t="dataTable" formula records degrade on rewrite — data-table interiors bake to static values #430, design-panel-directed): CellValue.Formula gained kind: FormulaKind (Normal | ArrayFormula | DataTable). <f t="array" ref=…> keeps its CSE marker and <f t="dataTable" ref dt2D dtr r1 r2/> interiors no longer bake to static constants on rewrite — read → model → dirty write on both backends, structural edits shift record payload geometry. This is the authoring substrate for feat: native two-variable Data Table authoring — the house sensitivity engine #419 (wave 21).
bytes-fidelity (ooxml: bytes-based reads have no SourceContext — preserved workbook children lost on read-bytes→write #412): bytes-read workbooks carry a SourceContext (SourceContent.OnDisk | InMemory); preserved workbook children and unknown parts survive bytes round-trips, law-tested byte-identical against path reads. API note: SourceContext.sourcePath: Path → content: SourceContent (no in-repo consumers; fromFile unchanged).
lint (lint: extensions — chartsheet/dialogsheet order tables, externalLink rels chain, [Content_Types] registration, SAX O(1) mode #413): chartsheet/dialogsheet child-order tables, externalLink r:id chain, [Content_Types] registration, O(1) SAX mode, and a new ref-out-of-bounds check that flags the structural: range shifts have no upper clamp — full-height ranges overflow past row 1048576 / col XFD (Excel refuses the file) #428 corruption class (previously reported clean on files Excel refuses).
flake (test: ExcelIOSpec 'single-pass is faster than auto-detect' is a wall-clock race — flakes under parallel load #414): the suite's only parallel-load flake replaced with structural single-pass assertions — 100/100 under saturated load where the original failed 3–4/25.

Gates (all green)

./mill __.compile 810/810, 0 errors, 0 non-exhaustive-match warnings
scalafmt CI form (reformatAll + checkFormatAll __.sources) — no drift
./mill -i __.test 1028/1028 tasks, 4,818 test cases
XL_ROUNDTRIP_MIN_SUCCESS=500 generative round-trip law green
./scripts/test-examples.sh + ./scripts/verify-skill-snippets.sh --local green

Field-repro replays (branch-built CLI, audit fixtures)

Full-height CF A1:XFD1048576 + insert-rows 2 2 → stays A1:XFD1048576; xl lint clean; openpyxl loads with zero warnings (previously the MultiCellRange warning)
Excel-stamped DV H5:H10 → H7:H12; autoFilter A5:D10 → A7:D12; Print_Area $A$1:$D$10 → $A$1:$D$12 (all previously inert)
B4 after insert: <f>A4*2</f><v>4</v> — single = in openpyxl, cache intact (previously ==A4*2, no <v>)
Array t="array" ref and dataTable t="dataTable" ref dt2D dtr r1 r2 + cached <v> survive an unrelated put (dataTable previously baked to a constant)
Over-max fixture (A144:XFD1048578) → xl lint reports [ref-out-of-bounds], exit 1

Integration notes

Two cross-cluster conflict resolutions (both semantic, both re-verified by the full suite): StructuralEditor.rewriteFormulas combines #427's equals-free/cache-preserving form with #430's FormulaKind threading (two cluster test pins updated to the merged semantics); XlsxWriter sheet-rels passthrough keeps #429's staleTableRels guard on #412's ctx.content accessor. One example updated for the 3-arg Formula pattern.

Follow-ups already filed: #435 (ca/aca attrs on Normal formulas; delete-band data-table degradation fidelity). Reviewer non-gating notes carried in the wave gaps list (delete-side cache staleness candidate, single-cell sqref emission shape).

Closes #412
Closes #413
Closes #414
Closes #427
Closes #428
Closes #429
Closes #430

🤖 Generated with Claude Code

…ched <v> The structural insert/delete rewrite printed every shifted formula with includeEquals = true and discarded its cached value. The writer serializes the stored expression verbatim, so '=A4*2' landed inside <f> (openpyxl reads it back as '==A4*2') and every formula cell displayed blank in cached-value viewers until a recalc. Print with includeEquals = false (the CF path and CLI putf convention; the reader's canonical model form is equals-free) and carry the cell's cachedValue through every successful shift — a successful shift means every reference survived, so the cache is still the Excel-valid display value. The None branch still degrades to #REF! and drops the cache. Refs #427 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…sertions The 'writeStream: single-pass is faster than auto-detect' test compared wall-clock elapsed times of the two write paths and asserted single-pass wins within 1.5x — a scheduling race, not a correctness property. Under 2x CPU oversubscription plus disk thrash it failed 3 of 25 iterations locally (e.g. single-pass 862ms vs auto-detect 412ms), matching the two field flakes on CI and in wave 18's parallel run. Replace it with two structural tests that pin what the timing was a proxy for, with no timing comparison: - 'single-pass with hint matches auto-detect output': with the hint equal to the true bounds, every ZIP entry (worksheet, shared strings, statics) is byte-identical between the two paths, and the data round-trips through readStream. - 'hinted single-pass streams to destination with dimension up-front': a mid-stream probe (evalTap at row 50 of 100) observes the destination file already materializing while the source is still producing — writeStreamImpl writes static parts and the worksheet header before the body segment pulls any row — and a superset hint (A1:C200 over A1:A100 data) is emitted verbatim, which no detection pass could do; auto-detect on the same data emits the detected A1:A100 as contrast. Both single-pass assertions fail when applied to the auto-detect path (discrimination verified), and each new test passed 50 of 50 iterations under the same load that broke the old one. Global-tmpdir assertions were deliberately avoided: xl-cli also creates xl-stream-* temp files in parallel test JVMs. Refs #414 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Insert-side shifts had no upper bound: a full-height range (CF sqref A1:XFD1048576, merges, DV) overflowed to row 1048578 / column XFE on insert and Excel refused the file as corrupt. Same gap in the formula shifter, which clamped only the lower bound. Extract the span algebra into Sheet.shiftSpan — ONE clamp site shared by merges, cf/dv envelopes, chart data refs (and the GH-429 consumers that follow): an insert pins the span end at Row.MaxIndex0/Column.MaxIndex0 and drops a span whose start passes the edge. FormulaShifter.shiftPos / shiftRangePos mirror it (SharedFormula.shiftedIndex bound): a single ref pushed past the edge -> None -> #REF!; a range end clamps, a range start past the edge voids the formula. Long intermediate math so pathological counts cannot overflow. Refs #428 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

XlsxReader.readFromBytes passed None where the path variant builds a SourceHandle, so a workbook read from bytes and written back dropped every preserved-but-unmodeled workbook child (externalReferences, workbookProtection, pivotCaches, ...) and every unknown part (externalLinks, pivotCache, media) — the #397 field symptom by other means. The archive is already resident on a bytes read, so there is no streaming excuse for the asymmetry: - SourceContent (xl-core): OnDisk(path) | InMemory(bytes), the physical origin a SourceContext preserves from. SourceContext.sourcePath: Path becomes content: SourceContent (fromFile keeps its signature; fromContent added). - SourceHandle becomes an enum; readFromBytesWithWarnings wraps a private clone of the caller's array (parse and context share the snapshot, so caller mutation can never desync them) and fingerprints it like a file. - XlsxWriter: withZipFile(path) generalizes to withSourceZip(SourceContent) — ZipFile random access on disk, ZipInputStream scan in memory — behind a SourceZip adapter consumed by every preservation read (structure, styles metadata, worksheets, SST, rels, media hashing, copyPreservedPart); copyVerbatim verifies the fingerprint over the bytes and writes them. - StyleIndex.fromWorkbookWithSource opens the archive from either source. Law (XlsxReaderPassthroughSpec): write(readFromBytes(bytes(F))) is byte-identical to write(read(F)) over a preservation-rich fixture (externalReferences + workbookProtection + pivotCaches + externalLink + pivotCacheDefinition + STORED media), both clean and after an identical edit; caller mutation after the read cannot corrupt the write. The GH-397 lint acceptance now covers the bytes path too. Refs #412 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…in, Content_Types registration, O(1) SAX mode, over-max ref detection Four GH-413 extensions plus the GH-428 corruption-class detector, all diagnostics-only (no reader/writer behavior changes): - CT_Chartsheet / CT_Dialogsheet canonical child-order tables; sheet-class parts (worksheet/chartsheet/dialogsheet) share one scan pipeline keyed by the sheet rel type, including root-content and r:id checks - externalLink parts: <externalBook r:id> resolved against the part's sibling .rels (UnresolvedRelId / WrongRelType), one more hop past the workbook-level externalReference check; new XmlUtil.relTypeExternalLinkPath - MissingContentType: every present-and-referenced part (internal rel target of any .rels lint walks, including _rels/.rels) needs an Override or an extension Default; absent [Content_Types].xml is a single finding - RefOutOfBounds: every ref=/sqref= token (dimension, mergeCells, conditionalFormatting, dataValidations, autoFilter, shared formulas, table parts) checked against row 1048576 / column XFD -- the GH-428 insert-overflow class Excel refuses as corrupt; boundary-max stays clean so post-clamp writer output lints clean (verified against both field repros: insert-rows -> A1:XFD1048578 and insert-cols -> A1:XFE1048576) - lintStream / lintStreamBytes: SAX scanning of sheet-class and table parts, O(1) memory in the row count; findings identical to DOM mode, pinned by a 16-fixture parity suite Refs #413 Refs #428 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

@ref

…bles, autoFilter on structural edits Excel stamps showInputMessage/showErrorMessage on virtually every validation it writes, so nearly all real-world DVs fell out of the typed model into Preserved — structurally inert, silently detaching from their data on every insert/delete. Print areas, tables, and the sheet-level autoFilter never shifted at all. - Widen the DV typed model (GH-375 -> GH-429): DvKind gains Custom, AnyValue, Bounded(whole/decimal/date/time/textLength x operator x formula2); Rules gains DvMessages (prompt/error text, errorStyle) with OOXML-default defaults; withPrompt/withError combinators and whole/decimal/date/time/textLength/custom/anyValue factories. - DataValidationCodec: widened whitelist + per-kind parse (operator only on bounded kinds; anything else still rides Preserved: xr:uid, imeMode, unknown types, corrupt sqref); emission in Excel stamp order with schema defaults omitted; message text through the new escapeXstringAttr (attr-value normalization eats raw LF/TAB/CR; law decodeXstring . escapeXstringAttr = id). planDvWrites unchanged — both CLEAN-gate sides use the same widened parser. - SqrefShift (new): byte-surgical root-sqref rewrite for Preserved DV/CF payloads under structural edits — layered all-or-nothing guards (start-tag region only, exactly one whitespace-preceded sqref, every token parseable), per-token identity fast-path, Excel-shaped printing for changed tokens; wired into shiftAxis for both Preserved cases. - Sheet.shiftAxis: pageSetup printArea/repeatRows shift through the GH-428 clamp (collapse clears the field); tables shift and retabulate on column edits (drop deleted columns, splice ColumnN-unique fresh columns, stamp their header cells; drop sub-minimum tables); autoFilter tri-state overlay (Ranged shifts, collapse -> Remove). - AutoFilterState (new) lift-and-overlay: reader lifts a parseable @ref; writer's mergeAutoFilterElem replaces only @ref (children verbatim, identity fast-path on equal refs), Remove strips, None passthrough — applied in both fromDomainWithMetadata branches. - StructuralEditor.rewriteDvFormulas: typed DV formulas shift through the same engine as CF/cell formulas (inline literals ride verbatim, fully-deleted sources -> #REF! text, cross-sheet sources tracked). - api exports: DvOperator, DvErrorStyle, DvBoundedType, DvMessages, AutoFilterState. Refs #429 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…rvive rewrites CellValue.Formula gains kind: FormulaKind (Normal | ArrayFormula | DataTable), the GH-430 binding representation. Records are read per cell (no group inference), re-emitted byte-exactly in CT_CellFormula schema order through the shared FormulaKindCodec on every writer (DOM, OoxmlCell SAX, DirectSaxEmitter, streaming), and survive DIRTY sheet regeneration — two-variable Data Table interiors no longer bake to static grids on put. - reader: WorksheetReader + SaxStreamingReader + SaxSingleCellReader recognize array/dataTable records leniently (junk attrs degrade visibly, never throw) - writers: dataTable emits a childless <f .../> (new SaxWriter.emptyElement; StAX writeEmptyElement override) with the cached <v> seed intact - evaluator: pinnedCache generalizes the GH-353 seam — DataTable caches pin (never parsed, never clobbered); cache write-backs use f.copy so kinds survive; DependencyGraph excludes DataTable nodes; StructuralEditor shifts ref/r1/r2 and degrades tearing edits (array->Normal text, dataTable->cached constant) - CLI: {=...} braces + additive JSON formulaKind; putf rejects top-level TABLE( pointing at GH-419; copy materializes dataTable cells and pastes array anchors as plain formulas - CellValue.dataTable smart constructor + FormulaKind api export prove the #419 authoring substrate (scripting-prelude probe) - new corpus fixture formula-records.xlsx rides every round-trip/parity law; property law: forAll genFormulaKind write->read = id on both backends Refs #430 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…rels Verifying the GH-429 collapse-drop path end-to-end exposed a writer hole: when a sheet's LAST table vanished (structural collapse or removeTable), fromDomainWithMetadata's tableParts fallback resurrected the preserved <tableParts> element and the copied sheet rels kept a table Relationship whose part no longer ships — our own reader refused the output ('Missing table file') and Excel would demand repair. Model-authoritative fix: the preserved tableParts element only rides while sheet.tables is non-empty, and a staleTableRels guard forces the rels merge path and drops table-type rels when the model has no tables left. Pinned by an integration test asserting no tableParts element, no table rels, no xl/tables/ part, and a clean re-read. Refs #429 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

The all-or-nothing guard in SqrefShift.shiftPayload checked xml.startsWith("<" + rootLabel), but every real reader-produced Preserved payload is canonicalized by CfCodec.preservedXml via XmlUtil.compact, which prepends '<?xml ...?>\n' — so the guard failed on every real file and DataValidation.Preserved / ConditionalFormat.Preserved entries stayed structurally inert on row/column edits (the exact GH-429 defect this rewrite claimed to fix). shiftPayload now skips one optional '<?...?>' prologue plus following whitespace, applies every existing guard to the element region behind it, and splices the prologue back verbatim on rewrite; a mangled prologue (no '?>') fails the root-label guard and rides unchanged like any other guard failure. Guard semantics stay all-or-nothing. Test realism repairs (each red on the prologue-blind guard, verified by refutation): - SqrefShiftSpec: reader-shaped shift (exact imeMode payload dumped from a field fixture), xmlns/xmlns:xr start-tag shift, identity (incl. full-column fixed point), refuse-matrix behind the prologue, mangled prologue, and all-drop -> None - DataValidationSpec / ConditionalFormatSpec / StructuralCfSpec flipped tests now use declaration-prefixed payloads and pin that the declaration survives the rewrite byte-verbatim - DataValidationCodecSpec commuting-square property derives payloads via CfCodec.preservedXml (the production shape) instead of Elem.toString, re-entering shifted payloads through a standalone parse like the emission path does - StructuralWritePreservationSpec: end-to-end pin on a hand-built foreign fixture (imeMode DV D5:D6, unknown-type DV E5:E6, xr:uid CF block B5:B6): read -> insertRows(1,2) -> Preserved sqrefs move to D7:D8/E7:E8/B7:B8 with the rest of each payload byte-intact -> write -> re-read clean Refs #429 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…-clock race

…e preservation, shift clamps, SqrefShift for preserved payloads

…ray/dataTable records Conflict resolution in StructuralEditor.rewriteFormulas: combined GH-427 semantics (includeEquals=false, cachedValue preserved on successful shift) with GH-430's FormulaKind threading (DataTable payload-geometry arm + shiftedArrayKind on the general arm) — formula-records branched pre-#427 and carried the old form.

…eContext Conflict resolution in XlsxWriter sheet-rels passthrough: kept GH-429's !staleTableRels guard AND GH-412's ctx.content accessor.

…detection

The cluster wrote these pins against its pre-#427 base (equals-prefixed re-print, cache discarded). Post-merge, shifted formulas re-print equals-free and keep their caches. Refs #430 #427

…430

Refs #412 #413 #414 #427 #428 #429 #430

claude · 2026-07-22T22:05:31Z

Review — Wave 19: structural integrity

I focused the review on the highest-risk surface area: the byte-surgical SqrefShift rewrite, the edge-clamp span algebra, the merged StructuralEditor.rewriteFormulas (the #427/#430 conflict resolution), and the bytes-fidelity SourceContent path. Overall this is high-quality, unusually well-documented work — the scaladoc on SqrefShift, shiftSpan, and FormulaKind explains why each guard exists, and the test-file-to-source ratio (49 test files touched) is excellent. The commit messages are exemplary root-cause narratives.

Strengths

Overflow-safe clamp algebra. Sheet.shiftSpan and FormulaShifter.shiftPos/shiftRangePos do the intermediate arithmetic in Long and only narrow after the axisMax comparison, so a pathological count/delta can't wrap. Verified the full-height CF case by hand: A1:XFD1048576 + insert-rows 2 2 → start 0 < at stays, end 1048575+2 clamps to 1048575, i.e. unchanged — matches the PR's field replay. One shared clamp site (shiftSpan) for merges/CF/DV/print-area/tables is the right factoring.
SqrefShift guards are genuinely all-or-nothing. The prologue-skip fix (commit 44246a6) is a real bug the earlier form would have silently no-op'd on every reader-produced payload — good catch, and the refutation-tested specs pin it. The reliance on scala.xml canonicality (> and " escaped everywhere, so the first > ends the start tag) is sound given the payload always comes from CfCodec.preservedXml; the guard degrades safely to unchanged on any non-canonical input.
Bytes-fidelity is memory-safe. readFromBytesWithWarnings wraps a private bytes.clone() before unsafeWrapArray, so caller mutation after the read can't desync parse-vs-context, and SourceContent.rawArray only exposes the backing array of that private snapshot. Symmetry with the path read is law-tested byte-identical.
FormulaKindCodec is lenient-total — junk t/ref/r1/r2 degrade to None/dropped attrs, never throws, never invents a range. DV message text is XML-escaped by the elemOrdered serializer on top of the xstring layer, so no attribute injection.

Findings / discussion

Data-table band-tear can silently drop cell content (StructuralEditor.scala:105). When a structural edit tears a DataTable interior, the cell degrades to cachedOpt.getOrElse(CellValue.Empty). If the record carried no cached <v>, the formula becomes Empty — the cell's content vanishes entirely on an insert/delete, with no warning surfaced to the caller. The degradation-to-cached-constant behavior is deliberate and documented (and del1/del2 fidelity is filed as ooxml: preserve ca/aca on plain formulas; set del1/del2 when a data-table input is deleted #435), but the no-cache → Empty corner is a quiet data loss. Consider emitting a warning/diagnostic (or a lint finding) when this path drops a formula, so it isn't invisible to a user doing a bulk row delete.
Documented staleness in SqrefShift / AutoFilterState — worth a user-facing note. Absolute refs in un-rewritten Preserved formulas go stale below an edit, and AutoFilterState interior column edits misalign filterColumn@colId. Both are correctly documented in the scaladoc and are strictly better than the prior full-inertness. No change requested — just flagging that these belong in docs/LIMITATIONS.md (already touched in this PR — good).
FormulaKindCodec.toAttrs fixed schema order vs. attribute sorting. The codec emits <f> attributes in CT_CellFormula schema order to reproduce Excel byte-for-byte, but CLAUDE.md notes XmlUtil.elem sorts attributes by name. Assuming the <f> emission bypasses the sorting path (the round-trip parity law would catch a mismatch), this is fine — I couldn't see a conflict, but it's the one spot where the "byte-for-byte" claim depends on emission not re-sorting. The parity suite covers it either way.

Verification note

I reviewed statically and did not independently re-run the suite; the PR reports 1028/1028 tasks / 4,818 cases green plus the generative round-trip law. The clamp and rewrite logic I traced by hand matches the field-repro assertions in the description.

Nice work — the #428 corruption class (A144:XFD1048578) closure and the equals-free/cache-preserving rewrite are meaningful correctness wins. My only substantive ask is surfacing finding #1 (silent no-cache data-table drop) rather than leaving it invisible.

The wave-19 docs described the integration-time cache-preservation policy; 8c021a9 correctly replaced it with invalidation (shrinking-delete aggregates and position-sensitive formulas make preserved caches silently wrong). Refs #427

arcaputo3 and others added 17 commits July 22, 2026 16:21

merge cluster flake (#414): structural writeStream tests replace wall…

b620ebd

…-clock race

merge cluster structural (#427 #428 #429): equals-free rewrite + cach…

18f8f54

…e preservation, shift clamps, SqrefShift for preserved payloads

merge cluster bytes-fidelity (#412): SourceContent bytes-backed Sourc…

1b37d15

…eContext Conflict resolution in XlsxWriter sheet-rels passthrough: kept GH-429's !staleTableRels guard AND GH-412's ctx.content accessor.

merge cluster lint (#413): five lint extensions incl. over-max range …

49679cc

…detection

test: reconcile FormulaRecordEvalSpec with merged GH-427 semantics

3eb7b5e

The cluster wrote these pins against its pre-#427 base (equals-prefixed re-print, cache discarded). Post-merge, shifted formulas re-print equals-free and keep their caches. Refs #430 #427

examples: 3-arg CellValue.Formula pattern (FormulaKind ripple). Refs #…

5c25b4e

…430

docs: wave 19 integration — CHANGELOG, roadmap record, test counts

c7ccb4a

Refs #412 #413 #414 #427 #428 #429 #430

arcaputo3 added 2 commits July 23, 2026 12:07

fix: address structural and formula record regressions

8c021a9

arcaputo3 merged commit aa6de9e into main Jul 23, 2026
4 checks passed

arcaputo3 deleted the wave19-structural branch July 23, 2026 17:18

This was referenced Jul 23, 2026

fix(cli): structural commands recalculate before writing #437

Merged

chore(release): Bump version to 0.16.0 #439

Merged

Navigation Menu

Uh oh!

Wave 19: structural integrity — shift clamps, cache/equals fidelity, formula-record + bytes preservation, lint teeth - #436

Wave 19: structural integrity — shift clamps, cache/equals fidelity, formula-record + bytes preservation, lint teeth#436
arcaputo3 merged 19 commits into
mainfrom
wave19-structural

arcaputo3 commented Jul 22, 2026

Uh oh!

claude Bot commented Jul 22, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Uh oh!

Conversation

arcaputo3 commented Jul 22, 2026

What's in it

Gates (all green)

Field-repro replays (branch-built CLI, audit fixtures)

Integration notes

Uh oh!

claude Bot commented Jul 22, 2026

Review — Wave 19: structural integrity

Strengths

Findings / discussion

Verification note

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant