v3.2.1 — Free canvas panning

v3.2.1 — Free canvas panning

• Release date: 13 October 2026.
• Base: v3.2.0.
• Scope: Recoverable canvas movement in ERD and SQL Flow.

What changed

• ERD canvas movement works in every direction, including right and down from
  the initial top/left position. Positive viewport offsets provide space before
  the diagram; negative movement uses an extended scroll area.
• ERD Fit recentres the complete schema, or a selected table and its declared
  neighbours. Cursor-anchored zoom accounts for the viewport offset. Panning
  moves the view independently of saved table positions.
• SQL Flow distinguishes a drag from a node click. Dragging the canvas or a
  node preserves the requested movement instead of selecting the node and
  reframing the diagram. Fit overview recovers the complete chart.
• Diagram controls remain interactive during panning. The user guide explains
  canvas movement, individual table movement, and Fit recovery.

Verification

An isolated Edge 155 walkthrough reproduced both problems before the fix.
Afterward, an 80-pixel right / 60-pixel downward drag moved both diagrams by
that amount; flow inspection stayed closed. ERD movement beyond the origin,
Fit recovery after a large negative pan, and cursor-anchored zoom were also
observed. The prior browser diagnosis used actual renderer pointer events.

Local TypeScript build, strict pilot, and lint (69 files) passed. The v3.2.1
fixture-corpus metrics snapshot was generated. A separate AI code reviewer
approved the signed viewport coordinates, zoom/Fit calculations, flow click
suppression, and guide changes.

The release requires green Correctness and Firefox smoke CI on its
complete release commit. Correctness covers generated-runtime freshness,
release-document consistency, local-file startup, golden/fuzz/UI/ERD/security
suites, current metrics, benchmark, and packaged-runtime smoke.

Runtime

The release includes sql-cartographer-v3.2.1.zip (4,069,229 bytes) and
SHA256SUMS.txt.

ZIP SHA-256:
530008c7aceba684dc3d1acd2d54244bc6cc18910e0e44c962857d95ee76ce80.

Download the GitHub release.

Correctness
and Firefox smoke
passed on release commit 2ec36db7451e200a934817c040892c05a65e7d48.

v3.2.0 — Clear diagrams and guided exploration

v3.2.0 — Clear diagrams and guided exploration

• Release date: 13 October 2026.
• Status: Browser review and local release qualification passed.
• Base: v3.1.0.
• Scope: More readable diagrams, inspectable evidence, actionable diagnostics, four curated lessons, and smaller presentation modules.

What changed

• Flow diagrams keep labels readable on arrival. Fit overview remains an
  explicit action; selecting a step highlights and frames its immediate
  modelled neighbours. Clearing focus restores the previous viewport.
• A flow inspector explains node class, provenance, available catalogue
  resolution, source evidence, and incoming/outgoing relationship kinds.
  Source and object navigation happen through explicit buttons.
• Flow and ERD findings share a text-safe diagnostic view with severity, stable
  code, scope, message, source location, and next-step hints for common codes.
• Three flow/query lessons cover a decision, temporary-table data flow, and a
  CTE pipeline. A three-table schema lesson connects declared foreign keys to
  a query plan and its row-repetition explanation. Guided and representative
  samples ask before replacing non-empty input.
• Flow diagram presentation and ERD card/inspector builders now have typed UI
  boundaries. The contributor guide follows a lesson from sample SQL through
  analysis and the graph to exported output.
• Downloaded SVGs omit temporary focus decoration while the on-screen view
  remains selected. Existing workspace, query, and layout formats are retained.
• Source navigation resolves the owning object, report XML, or catalogue input.
  Stale actions preserve edited input, and clearing SQL clears old diagnostics.
• At 200% browser zoom, stacked controls wrap and scroll with the page; keyboard
  focus keeps the flow inspector and zoom controls inside the canvas.

Experience review

The maintainer review and its stated limits are recorded in
the v3.2.0 review sheet. It includes the actual Edge
200% zoom walkthrough, keyboard interactions, light/dark captures, and fixes.
No timed first-time-user study or independent human walkthrough is claimed.
Representative screenshots:

Verification

Gate	Result
npm run lint	Pass — 69 files
npm run typecheck	Pass
npx tsc --project tsconfig.strict.json	Pass
npm run build	Pass
npm test	Pass — golden 222/222; fuzz 400 cases; UI 67/67; ERD UI 74/74; security 7/7
npm run test:file	Pass — local-file runtime loaded
npm run metrics:write / npm run metrics	Pass — v3.2.0 snapshot generated and current
npm run package:smoke	Pass — archive structure and local-file smoke
npm run benchmark	Pass — 0 overlaps/crossings across recorded graph fixtures
Actual Edge 200% zoom walkthrough	Pass — flow/report/schema evidence, keyboard source/query actions, both themes, no page overflow
Separate AI code review	Approved — source ownership, stale input, responsive fixes, and release-doc gate

The local runtime archive is .release/sql-cartographer-v3.2.0.zip; its SHA-256
is 3ae3296373e9f2c2772a5f69da4aec2d422a77b94629af93d6c77753b8d68a91.
The 4,067,501-byte archive and .release/SHA256SUMS.txt accompany the
GitHub release.

The runtime archive and SHA-256 checksum are built by npm run package:runtime.
The package smoke verifies that archive before publication. The visual review
limits are listed in the linked review sheet.

v3.1.0 — Documentation, diagrams, and maintainability

SQL Cartographer v3.1.0 — Documentation, diagrams, and maintainability

A quality pass over the code and the documentation. The product now explains
itself: the docs carry real Mermaid diagrams and screenshots of both pages, the
architecture is described against the code that actually exists, and the
contributor path has one home. Two correctness defects found during the pass
are fixed.

No user-data format changed. Nothing saved by v3.0.0 is affected.

Highlights

• Diagrams in the documentation. The project has exported Mermaid since
  v1.0 and not one doc rendered a diagram. There are now Mermaid figures for
  the analysis pipeline, the module dependency graph, the ERD layout pipeline,
  bridge tables, self joins, the two-paths tie-break, the report → dataset →
  object → column chain, and the node/edge vocabulary. All eight parse under
  securityLevel: 'strict' with the vendored renderer.
• Screenshots of both pages. index.html and erd.html — including the
  query builder mid-plan — are pictured in the README and the user guide. For a
  visualisation tool, this was the largest "can I judge this before running it"
  gap.
• A worked example. The user guide now walks one procedure from source SQL
  through the rendered diagram to the exported Mermaid and its provenance
  comment, using the product's real output rather than an invented one.
• A wordmark defect fixed. v3.0.0 renamed the project and claimed one
  consistent wordmark, but both pages still shipped proc>flow because the
  markup splits it across a <span class="caret">. A CI guard now catches that
  exact form.

Fixed

• ROLLBACK TO <savepoint> misread its target. tsqlTransactionAction
  never skipped the TO keyword, so ROLLBACK TO my_sp reported the savepoint
  as TO. The informative "roll back to savepoint X" label could not fire, and
  the name would have been registered wrongly had the save path seen it. Found
  by writing direct unit fixtures for the transaction model; those fixtures are
  the regression test.
• A failed column-flow analysis vanished silently. Two guards in
  analyse caught a failure and returned no column flow with no diagnostic,
  contradicting the contract that unknown constructs stay explicit. Both now
  push a document-scoped column_flow_analysis_error, matching what the report
  pipeline already did. The column-flow model survives a graph-layout failure;
  only the rendered graph is dropped, and that is now said out loud.

Documentation

• docs/README.md — a documentation map with a learning path for users,
  contributors, and reviewers.
• CONTRIBUTING.md — the non-negotiables, fixtures-before-fixes, and the
  pre-PR checklist, consolidated from three places that had drifted. ADR
  template and index in docs/decisions/.
• docs/GLOSSARY.md — the vocabulary the docs and the UI share (opaque region,
  bridge table, taught join, token attribution, declared evidence, and friends),
  which until now was used before definition.
• .github/SECURITY.md — points GitHub's security tab at the canonical policy
  in docs/SECURITY.md.
• docs/ARCHITECTURE.md refreshed against src/: seven modules were missing,
  one attribution was stale, and the documented one-way dependency chain was
  not reflected. Now includes the module load-order contract and the graph
  vocabulary table.
• TOCs added to the user guide and the query builder. 21 broken relative links
  in docs/archive/ repaired (they broke when those notes moved into
  archive/). README's roadmap link no longer contradicts the roadmap's own
  "historical" banner.

Maintainability

• The T-SQL transaction model is now testable on its own. 15 pure functions
  moved from src/ir.ts into src/dialects-state.ts, alongside the
  TSQL_XACT_* bitmask constants they operate on. ir.ts no longer contains
  the bitmask algebra at all.
• emitOne is a dispatch, not a 400-line switch. Its 16 node-kind cases
  are now named emitters (emitTry, emitIteration, emitHandler, …), each
  typed to its own AST node. ir.ts is 1532 → 1057 lines and the emitted graph
  is unchanged, verified by 222 golden assertions.
• 54 diagnostic codes are compiler-enforced. Diagnostic.code is now a
  closed DiagnosticCode union instead of string, so a typo is a compile
  error rather than a message that never matches a test.
• Deduplicated: twelve localStorage wrappers in src/workspace.ts
  collapsed to four primitives; three copies of the same span predicate in the
  golden tests collapsed to one; the duplicated outer-parenthesis stripper in
  src/dialects-state.ts is one helper.
• Renamed the confusables: queryBuildGraph → queryBuildJoinGraph
  (distinct from lineage.ts's buildQueryGraph), the O/o parameter pair
  in src/columnflow.ts, a variable named any, and the version-suffixed test
  helpers.
• Bundled ERD sample DDL moved from src/erd.ts to src/samples.ts, whose
  stated purpose is exactly that. Removed the production-dead layoutDrawio
  wrapper.

Testing

• New suite tests/xact-state.ts: 58 direct fixtures for the transaction
  algebra, which previously had zero direct coverage — depth-range
  intersection, the depth ↔ state bit-set conversions, the state labels,
  transaction action parsing (including BEGIN DISTRIBUTED TRAN [WORK] and
  variable savepoint targets), action validity per state, and FlowContext
  parent-chain inheritance. It gates the golden page.
• Two new fixtures in tests/column-flow.ts require that a column-flow or
  column-graph failure is reported rather than swallowed.
• golden 221 → 222 cases. The fixture-corpus snapshot is
  metrics-v3.1.0.json, with columnFlow 15 → 17 and
  every tracked rate still 1 — attribution, semantic edge coverage,
  provenance, export parity, export traceability, layout budgets, and the
  adversarial semantic assertion rate.

Verification

npm run lint, npm run typecheck, npm run build, npm test
(728 cases: golden 222, fuzz 400, ui 32, erd-ui 68, security 6),
npm run test:file, npm run metrics, npm run benchmark, and
npm run package:smoke all pass. Committed dist/ is current. The runtime
remains local-only: no network call, and localStorage is still touched only
by src/workspace.ts.

Known gaps

Deliberately out of scope for this release, and recorded so they are not lost:

• src/ui/erd-page.ts is still 1422 lines (cards, layout, zoom, routing, and
  event wiring in one closure). The clean split already used by
  src/ui/erd-query-view.ts is the template, but extracting it is a
  behaviour-sensitive change to the ERD interaction layer and wants its own
  release with its own fixtures.
• The type-strictness pilot still covers only the extracted modules. Extending
  tsconfig.strict.json is incremental work that is worth doing file by file.
• Test-side duplication of the export-manifest assertion block across
  tests/parity.ts, tests/report-graph.ts, and tests/column-flow.ts remains.

See the changelog for the condensed history.

v3.0.0 — Rename and identifier reset

SQL Cartographer v3.0.0 — Rename and identifier reset

The project was named ProcFlow until v2.7.0. From v3.0.0 it is
SQL Cartographer, published at
ikelaiah/sql-cartographer.
This is a breaking release: the rename reaches user-data-facing identifiers,
so saved work from v2.7.0 and earlier does not load. There is no migration
path.

Highlights

• New name everywhere that matters. The GitHub repository, the npm package
  (sql-cartographer), the product name in prose, the release archive, the
  logo, and the CI and packaging scripts all follow the new name. The old
  ikelaiah/ProcFlow repository URL redirects automatically.
• One consistent wordmark. The logo now reads sql>cartographer, the
  exported Mermaid provenance header reads %% sql-cartographer provenance,
  and generated SQL headers are prefixed sql-cartographer.
• Failure output that names the failure. npm test previously printed only
  erd-ui: 66/68 tests passed, which is enough to know something broke and
  not enough to know what. Each failing case is now printed with its name and
  detail.
• Hardened ERD interaction polls. The erd-ui polls that wait for
  requestAnimationFrame-driven overlay drawing allow 10s instead of 3s, so a
  contended CI runner does not drop cases a desktop browser passes.

Breaking: user-data identifiers

Anything saved before this release is stored under the old identifiers and
will not load. Export what you want to keep before upgrading.

Kind	v2.7.0 and earlier	v3.0.0
Browser storage keys	procflow.workspace, procflow.erd.layout, procflow.erd.query	sql-cartographer.workspace, sql-cartographer.erd.layout, sql-cartographer.erd.query
Saved-query file format	procflow-erd-query	sql-cartographer-erd-query
ERD-layout file format	procflow-erd-layout	sql-cartographer-erd-layout
draw.io XML attributes	data-procflow, data-procflow-kind	data-sql-cartographer, data-sql-cartographer-kind
draw.io diagram id	procflow-page	sql-cartographer-page
Default download names	procflow-report, procflow-estate, procflow-workspace.json, procflow-query-<fp>	sql-cartographer-report, sql-cartographer-estate, sql-cartographer-workspace.json, sql-cartographer-query-<fp>
Release archive	procflow-vX.Y.Z.zip	sql-cartographer-vX.Y.Z.zip

Before upgrading:

1. Saved queries — use Query → Export to write a procflow-query-*.json
   file. After upgrading, Query → Import accepts the same file: the importer
   recognises a procflow-erd-query payload and re-saves it under the new
   format, so only the file needs to survive the upgrade.
2. ERD layouts — use Export to write a procflow-erd-layout-*.json file.
   Layout files are read by format tag rather than by filename, so an old
   export imports cleanly and is rewritten under the new tag.
3. Workspaces — a procflow-workspace.json export is not readable by
   v3.0.0. The workspace format is version-checked and rejected rather than
   half-imported. Rebuild the workspace after upgrading, or stay on v2.7.0
   until you no longer need it.

Browser storage is not migrated. The old procflow.* keys remain in the
browser but are never read; clear them manually if you want the space back.

Accuracy and limits

• No analysis, export, or query-builder semantics changed. The rename is a
  mechanical identifier substitution: every golden, fuzz, adversarial, and
  security fixture passes with the same expectations as v2.7.0, and the
  fixture-corpus metric numbers are identical.
• The docs/metrics-v*.json snapshots for v2.7.0 and earlier keep the
  generator label that release produced, except metrics-v2.7.0.json which
  was regenerated with the new label so the correctness workflow's snapshot
  currency check stays meaningful. The metric numbers are unchanged either way.
• Historical documents keep the old name on purpose: docs/archive/, the
  canonical release notes for v1.14.0 through v2.6.0, and the CHANGELOG.md
  entries for those releases describe what shipped under it.
• Browser automation depends on a locally installed Chromium/Edge/Chrome;
  npm test honours CHROME_PATH and fails with a clear message otherwise.

Verification

The final local verification recorded for this release includes:

• npm run lint (Biome) passing with 0 errors;
• npm run typecheck and npm run build passing;
• npm test (scripts/test-all.mjs) passing all five browser suites over a
  local static server: golden 221/221, fuzz 400 deterministic mutation cases,
  ui 32/32, erd-ui 68/68, security 6/6;
• npm run test:file (local-file startup smoke) passing;
• npm run metrics confirming the v3.0.0 fixture-corpus snapshot is current;
• npm run benchmark completing the repeatable scalability run;
• npm run package:smoke passing archive structure and local-file startup on
  sql-cartographer-v3.0.0.zip;
• CI runs the same gate plus Firefox smoke and release-consistency checks.

Upgrade and distribution

Download sql-cartographer-v3.0.0.zip from the GitHub release or clone
ikelaiah/sql-cartographer.
The runtime footprint is unchanged apart from the identifier rename: the same
index.html, styles.css, and dist/ modules. Open index.html (or
erd.html) in a current Chromium, Firefox, or Edge browser.

If you are linking to the old repository, update the URL — the redirect works
for clones and releases, but GitHub Pages URLs do not redirect and the live
demo is now at https://ikelaiah.github.io/sql-cartographer/.

v2.7.0 — Maintainability pass

A code-health release: no new user-facing features, but a stronger codebase
underneath the same UI. The analysis core is now a clean one-way dependency
chain, duplication is removed, dead code is gone, linting exists, and
contributors can run every browser suite with one command.

Highlights

• No dependency cycles in the analysis core. Shared token/AST helpers moved
  to src/token-utils.ts and the PL/pgSQL + T-SQL transaction rules to
  src/dialects-state.ts. token-utils ← lineage ← columns ←
  columnflow is one-way, and exporters no longer reaches back into
  ir.ts. Each module can now be read and tested on its own.
• ir.ts shrinks from 1966 to 1466 lines. The graph builder keeps its
  responsibilities, but the pure helpers and dialect state live where a reader
  can find them.
• One CTE splitter. splitCTEs now carries the explicit column list
  (WITH r(n) AS …) and body spans that previously lived in a near-duplicate
  splitCtesColumn; lineage and column flow share one parser.
• One filter kernel. The dependency and report graph filters shared a
  copy-pasted neighbour/clone block; both now call graphNarrowByFocus and
  graphFilteredClone, so the no-mutation and one-hop rules cannot drift.
• Exports never mutate the analysis. toMermaid works on shallow node
  copies, so a cached graph is safe to reuse; a regression test locks this in.
• npm test runs the whole browser suite. scripts/test-all.mjs starts a
  local static server and drives golden, fuzz, ui, erd-ui, and security
  through headless Chromium. The suites also publish named result globals so
  the runner does not scrape DOM state.
• Linting and CI tightening. Biome (npm run lint) gates correctness and
  suspicious rules; CI runs lint before typecheck. noUnusedVariables is
  intentionally off because the global-script architecture makes top-level
  declarations look unused to a per-file linter.
• Tests are no longer order-dependent. The UI suites previously raced the
  off-screen iframe: on a fast machine the iframe could load before the test
  attached its listener, leaving the suite stuck at "Running…". They now run
  either way.
• Windows-friendly metrics. npm run metrics normalises line endings, so
  the snapshot check no longer reports a false staleness on CRLF checkouts.
• UI polish. Tablists follow the ARIA tabs keyboard pattern (arrows,
  Home/End), and the app supports prefers-color-scheme: light with a
  legible light palette. Command-menu panels carry role="group" so their
  labels are exposed to assistive technology.
• Docs that point forward, not backward. 33 superseded RELEASE_NOTE_*,
  PR_NOTE_*, and historical planning documents moved to docs/archive/
  with an archive README; ROADMAP.md, tasks/, and BENCHMARKS.md are
  clearly marked historical. docs/RELEASES.md now indexes every canonical
  release note. The stale trust-model version and the port mismatch in
  docs/DEVELOPMENT.md are corrected.

Accuracy and limits

• No analysis, export, or query-builder semantics changed. Every golden,
  fuzz, adversarial, and security fixture passes unchanged; the metrics
  snapshot is byte-identical to v2.6.0's numbers.
• The metrics-v1.6.0…v1.12.0 snapshots referenced by archived v1.x notes
  do not exist in the repository; the notes stay archived as historical
  records and their broken references are left as-is.
• Browser automation depends on a locally installed Chromium/Edge/Chrome;
  npm test honours CHROME_PATH and fails with a clear message otherwise.

Verification

The final local verification recorded for this release includes:

• npm run lint (Biome) passing with 0 errors;
• npm run typecheck and npm run build passing;
• npm test (scripts/test-all.mjs) passing all five browser suites over a
  local static server: golden 221/221, fuzz 400 deterministic mutation cases,
  ui 32/32, erd-ui 68/68, security 6/6 — including the new export
  non-mutation and Mermaid strict-security regression tests;
• npm run test:file (local-file startup smoke) passing;
• npm run metrics confirming the fixture-corpus snapshot is current;
• npm run package:smoke passing archive structure and local-file startup;
• CI runs the same gate plus Firefox smoke and release-consistency checks.

Upgrade and distribution

Download procflow-v2.7.0.zip from the GitHub release or clone the
repository. Nothing changes in the runtime footprint beyond the new
token-utils.js, dialects-state.js, and samples.js modules; the
package allowlist and index.html load order are updated accordingly.
Open index.html (or erd.html) in a current Chromium, Firefox, or Edge
browser.

v2.6.0 — Aggregates and grouping

ProcFlow v2.6.0 — Aggregates and grouping

Released as the aggregate milestone. The query builder now answers grouped
questions — orders per customer, totals per product — without leaving the
declared-evidence model or hand-writing GROUP BY.

Highlights

• Aggregate any picked column. Each chip carries a selector: —,
  COUNT, COUNT DISTINCT, SUM, AVG, MIN, MAX.
• GROUP BY derives itself. Every remaining picked column becomes the
  group list in pick order; when all picks are aggregated, GROUP BY is
  omitted and the query returns one row. No second multi-select, no invalid
  states.
• Aggregates read like SQL. Deterministic aliases
  (count_orderheader_orderid), sorting an aggregated column orders by the
  expression (ORDER BY COUNT(...) DESC), and the provenance header records
  Grouped by: and Aggregates: lines.
• DISTINCT yields to grouping. While an aggregate is set the toggle is
  disabled and the SQL omits the clause, with a header note explaining why.
• Fan-out honesty. A plan tip warns that SUM/AVG over a one-to-many
  join can multiply unless the detail rows are pre-aggregated.
• Versioned files stay compatible. Query files move to version 2;
  version-1 files from v2.5.0 load and migrate forward with an empty aggregate
  map, and newer versions are rejected with a diagnostic. Aggregates are
  saved, exported, restored, and pruned with their picks.
• Documentation. ADR-003 records the derived-GROUP BY model and the
  rejected alternatives (HAVING, explicit grouping, COUNT(*), window
  functions); docs/QUERY_BUILDER.md gains an aggregates section.

Accuracy and limits

• No analysis or join semantics changed. Joins still use declared foreign-key
  evidence only; taught joins and self-join copies remain labelled not
  declared; the generated SQL stays read-only.
• No HAVING, no window functions, and no COUNT(*): COUNT(column) ignores
  NULLs, so count a NOT NULL key for row counts. GROUP BY covers picked
  columns only.
• Unknown aggregate functions are ignored, never guessed.

Verification

The final local verification recorded for this release includes:

• golden browser assertions on tests/index.html passing (221/221), including
  the query suite (55 records) with aggregate rendering, all-aggregate
  GROUP BY omission, aggregate ordering, DISTINCT suppression and header
  notes, unknown-function tolerance, version-1 file migration, and storage and
  pruning coverage;
• the ERD query UI suite passing 68/68 over HTTP, including the aggregate
  selector, function rendering, GROUP BY, DISTINCT disabling, the plan
  tip, and reset;
• the Firefox smoke page passing 10/10 critical paths;
• npm run typecheck, npm run build, npm run test:file, npm run metrics,
  npm run package:smoke, and npm run benchmark passing;
• runtime archive structure and local-file startup smoke passing
  (procflow-v2.6.0.zip).

Upgrade and distribution

The runtime is self-contained. Use npm run package:runtime to produce the
versioned ZIP and verify .release/SHA256SUMS.txt before distributing it.
Layouts, workspaces, and version-1 query files saved by earlier releases remain
readable.

v2.5.0 — Saved queries and file export

ProcFlow v2.5.0 — Saved queries and file export

Released as the query-persistence milestone. The ERD query builder no longer
loses work when the tab closes, and the generated SQL can leave the browser as
a file. Everything remains local: one explicit saved slot per browser, plus
versioned files you own.

Highlights

• Save to this browser, Restore saved, Forget saved. One query is stored
  under procflow.erd.query with explicit actions only — nothing is written
  automatically, matching the ERD layout and workspace rules. Activating query
  mode with a saved query and no picks points at Restore.
• Versioned query files. Export query file writes a
  procflow-erd-query v1 JSON payload carrying the selections, taught joins,
  self-join copy routing, options, and a schema fingerprint. Import query
  file reads it back.
• Download .sql. The statement is written as a file and respects the
  Comments toggle. The optional name field drives filenames
  (orders-by-customer.json / .sql); without it the schema fingerprint is
  used.
• Prune-and-report restores. References that no longer exist — tables,
  columns, taught joins, sorts, cross/excluded ids, path choices — are dropped
  and counted, and a changed schema is named in the status line. Restores never
  fail because the schema moved on. Foreign formats, future versions, and
  malformed JSON are rejected with diagnostics.
• Pure store module. src/query-store.ts owns serialization, validation,
  and pruning with no DOM or storage access; schemaFingerprint moved to
  src/schema.ts and is shared with ERD layout files. Browser storage stays in
  the CI-guarded src/workspace.ts.
• Documentation. ADR-002 records the persistence decisions and rejected
  alternatives; docs/QUERY_BUILDER.md gains the file format and drift
  semantics; ARCHITECTURE adds invariant 7.

Accuracy and limits

• No analysis or join semantics changed. Joins still use declared foreign-key
  evidence only; taught joins and self-join copies remain labelled not
  declared; the generated SQL stays read-only.
• One saved query per browser by design; keep multiple analyses as files. See
  ADR-002 for why a named library was deferred.
• Saved queries and ERD layouts are independent storage keys; forgetting one
  never drops the other.

Verification

The final local verification recorded for this release includes:

• golden browser assertions on tests/index.html passing (221/221), including
  the query-store records for deterministic round-trip, three rejection codes,
  unreadable-entry tolerance, stale pruning with reasons, fingerprint
  stability, export naming, and the opt-in storage round-trip;
• the ERD query UI suite passing 62/62 over HTTP (query mode, picking, join and
  problem cards, teaching, self joins, options, highlighting, resize, Find,
  only-used filter, export/import/download, and browser save/restore/forget);
• the Firefox smoke page passing 10/10 critical paths;
• npm run typecheck, npm run build, npm run test:file, npm run metrics,
  npm run package:smoke, and npm run benchmark passing;
• runtime archive structure and local-file startup smoke passing
  (procflow-v2.5.0.zip).

Upgrade and distribution

The runtime is self-contained. Use npm run package:runtime to produce the
versioned ZIP and verify .release/SHA256SUMS.txt before distributing it.
Layouts and layout files saved by earlier releases remain readable; saved
queries are a new key and do not affect them.

v2.4.1 — Query builder hardening and release discipline

ProcFlow v2.4.1 — Query builder hardening and release discipline

Released as the first patch after the v2.4.0 query builder. No analysis
semantics changed: this release hardens the builder, commits its browser
suite, fixes an overlay redraw defect, and makes release documentation
verifiable in CI.

Highlights

• Query engine hardening. Shortest-path enumeration is iterative, so a
  1,000-table declared chain cannot exhaust the call stack. A scale fixture
  asserts 999 joins, 998 bridges, and deterministic plan and SQL output.
• Plan-edge highlighting fix. The overlay redraw after a pick was deferred
  to requestAnimationFrame, which can be starved in headless and off-screen
  contexts; only a pre-plan resize draw could run, leaving every edge dimmed
  even though the plan was correct. Query-state changes now draw synchronously;
  animation frames still coalesce drag redraws.
• Panel corrections. Excluding a problem table now prunes its sorts (not
  just its picks); the clipboard fallback is shared by both pages; the dropdown
  teach flow explains that the same table on both sides creates a self join.
• Panel structure. Pure DOM builders moved to src/ui/erd-query-view.ts
  (panel 1,145 → 812 lines; view 376 lines) with no behavior change.
• Accessibility and keyboard. One concise status announcement replaces
  live-announcing the whole SQL and plan; the floating window is a labelled
  dialog; Q toggles query mode; Escape cancels teaching without clearing
  the diagram selection.
• Committed browser suite. tests/erd-ui.html drives erd.html with 49
  interaction checks (query mode, picking, join and problem cards, teaching,
  self joins, options, highlighting, resize, Find, only-used filter), and the
  Firefox smoke now loads both pages.
• Release discipline. package-lock.json tracks package.json; the
  changelog, README, release index, and lockfile version are guarded in CI;
  docs/RELEASES.md documents the verified tag-and-asset flow.
• Documentation. docs/QUERY_BUILDER.md deep dive, ADR-001 for the
  declared-evidence query model, and a browser-suite cookbook plus module
  header convention in docs/DEVELOPMENT.md.
• CI actions updated (actions/checkout, actions/setup-node).

Accuracy and limits

• No analysis semantics changed. Joins still use declared foreign-key evidence
  only; taught joins and self-join copies remain labelled not declared in the
  plan and SQL; the generated SQL stays read-only.
• Query picks and options remain session state; only ERD layouts persist
  (opt-in).

Verification

The final local verification recorded for this release includes:

• golden browser assertions on tests/index.html passing (221/221), including
  the query-builder engine suite with the 1,000-table scale fixture;
• the ERD query UI suite passing (49/49 over HTTP), and the Firefox smoke page
  passing 10/10 critical paths;
• npm run typecheck, npm run build, npm run test:file, npm run metrics,
  npm run package:smoke, and npm run benchmark passing;
• runtime archive structure and local-file startup smoke passing
  (procflow-v2.4.1.zip).

Upgrade and distribution

The runtime is self-contained. Use npm run package:runtime to produce the
versioned ZIP and verify .release/SHA256SUMS.txt before distributing it.
Layouts and layout files saved by earlier releases remain readable.

v2.4.0 — Query builder on the ERD

ProcFlow v2.4.0 — Query builder on the ERD

Released as the ERD query-builder milestone. The declared-constraint model now
produces runnable SQL: pick columns on the diagram and ProcFlow works out the
joins the DDL actually declares, shows the plan, and hands you the statement to
run in your database client.

Highlights

• Query builder on the diagram. A toolbar toggle turns every column row
  into a checkbox and opens a floating window that can be dragged, resized
  (corner grip or arrow keys), and collapsed. Picked rows are highlighted,
  cards carry a picked-count badge, and the floating window can be closed and
  reopened with the picks intact.
• Declared-evidence joins only. Shortest paths over declared foreign keys
  connect the picked tables, preferring exact resolutions over name-matched
  ones. Bridge tables are included, tagged, and never selected. Optional and
  parent→child traversals use LEFT by default in the preserve policy; every
  join has an INNER/LEFT switch and a plain-language note about direction,
  optionality, and row multiplication.
• Equal-cost paths are a choice, not a guess. When two declared paths reach
  the same table (for example CreatedBy and ApprovedBy), the plan shows
  both paths and the first declaration is used until you pick another.
• Unjoinable picks are explicit. Tables with no declared path get a coral
  ring and a problem card. Resolve by teaching the join (column pair, typed
  predicate, or click two columns on the diagram), opting into a CROSS JOIN,
  or leaving the columns out — the provenance header names anything omitted.
  ProcFlow never silently produces a cartesian product.
• Self joins. Clicking two columns of the same table creates a second
  aliased copy; tick which picked columns read from that copy, and the header
  documents the alias as second copy, self join.
• Query ergonomics. Dialect-aware DISTINCT, a row cap (TOP n / LIMIT n
  / FETCH FIRST n ROWS ONLY), and per-column sort cycling (ascending →
  descending → off) on the picked-column chips. A tip offers DISTINCT when a
  one-to-many join can repeat rows.
• Readable generated SQL. Dependency-free syntax highlighting (keywords,
  comments, strings, identifiers, numbers, punctuation) where the concatenated
  token text is byte-identical to the emitted SQL; Copy always copies the raw
  text. Dialect quoting covers T-SQL, PostgreSQL, DB2, and SQLite.
• Diagram and plan stay one surface. Plan edges draw at full strength while
  unrelated edges dim; hovering a join card lights its exact edge and both
  endpoint cards; Only used tables narrows the canvas to the query plus any
  unresolved tables; Find now matches columns and highlights the exact
  rows; hover and teaching cursors make the diagram the picker.

Accuracy and limits

• Joins use declared PRIMARY KEY / UNIQUE / FOREIGN KEY evidence only.
  Column names never imply a relationship.
• Taught joins, self-join aliases, and CROSS JOIN opt-ins are labelled as not
  declared in the plan and in the SQL header; ProcFlow states that it cannot
  verify them.
• Name-matched foreign-key targets remain usable with a warning; external
  (unresolved) targets cannot be picked and are explained in the picker notes.
• The generated SQL is read-only and derived. Edit it where you run it.
• Query picks and options are session state; only ERD layouts persist (opt-in).

Verification

The final local verification recorded for this release includes:

• npm run typecheck, npm run build, npm run test:file, and
  npm run metrics passing;
• golden browser assertions on tests/index.html passing (221/221), including
  the v2.4.0 query-builder suite (39/39: join graph and pathfinding, ambiguity
  alternatives, bridges, preserve/strict policy, taught joins, self joins,
  disconnected selections, dialect quoting, ergonomics options, highlighting
  token round-trip, and deterministic output);
• a scripted headless UI pass over erd.html passing 67/67 interaction checks
  (query mode, on-card picking, join and problem cards, teaching by clicks,
  self-join copy routing, resize and keyboard resize, highlighting, chips,
  Find and the only-used filter);
• runtime archive structure and local-file startup smoke passing
  (procflow-v2.4.0.zip).

Upgrade and distribution

The runtime is self-contained. Use npm run package:runtime to produce the
versioned ZIP and verify .release/SHA256SUMS.txt before distributing it.
Layouts and layout files saved by earlier releases remain readable; query
builder state is not persisted between sessions.

v2.3.0 — Readable estate layouts

ProcFlow v2.3.0 — Readable estate layouts

Released as the v2.3.0 layout-quality milestone. The ERD page now arranges a
whole estate so relationships are traceable, not just placed.

Highlights

• Auto arrange on first parse. A freshly parsed schema starts in the
  deterministic layered layout. Reset to declaration order is one click
  away, and any drag, restore, or import keeps manual positions.
• Port-anchored orthogonal routing. Edges leave the referenced card at the
  row of the foreign-key column and enter the referencing card at its column
  row. Paths are orthogonal, travel through the card-free gutters between
  layout columns, and use band corridors plus an outer vertical channel for
  edges that span multiple bands. Gutters are clamped under compact spacing so
  a path never enters a neighbor column.
• Card-safe by construction. Structured routing is used whenever the layout
  came from Auto arrange or the declaration-order default; manual layouts fall
  back to the best-effort orthogonal router. Verified zero path/card
  intrusions on the DB2 sample, a 300-table chain, and a sampled 800-table
  chain.
• Stronger crossing reduction. Median sweeps alternate with a monotonic
  adjacent-transpose pass that only accepts swaps which reduce crossings;
  crossing counts use inversion counting and are reported in the Layout status
  line (for example Auto-arranged 12 entities · 0 crossings · LR).
• Direction and spacing. Layout → Direction supports Auto / Left→right /
  Top→bottom; Auto keeps whichever direction fits a 16:9 viewport better.
  Layout → Spacing adds Compact / Normal / Roomy presets that scale all
  gaps deterministically (DB2 sample: 1780×656 compact vs 2114×824 roomy).
• Pinning. A diamond in each card header pins its position; Auto arrange
  keeps pinned cards exactly and shifts unpinned cards deterministically to
  avoid overlaps. Unpin all clears them; Reset clears pins with the
  layout.
• Downstream banding. Isolated tables and views without declared sources
  move to their own downstream band instead of crowding the root column.

Accuracy and limits

• Auto arrange remains a deterministic near-optimal layout, not a
  minimum-crossing guarantee; arbitrary schema graphs can still cross.
• Structured (card-safe) routing applies to left-to-right layouts; top-to-bottom
  layouts use the best-effort orthogonal router.
• Hub fan-outs use staggered gutters that read as a bundle; a single shared
  trunk with branch stubs is still future work.
• Layouts are UI state and do not affect the SQL analysis, Mermaid export, or
  the workspace snapshot.

Verification

The final local verification recorded for this release includes:

• npm run typecheck, npm run build, npm run test:file, and
  npm run metrics passing;
• golden browser assertions on tests/index.html passing (220/220), including
  the v2.3.0 fixtures for orientation, crossings, pinning, spacing presets,
  banding, and a 300-table overlap check (35/35 schema suite);
• 31/31 UI interaction checks and the hostile-input security suite passing;
• ERD smoke checks: DB2 sample 5/5 orthogonal routes with 0 intrusions; a
  300-table chain 300/300 routes with 0 intrusions; a sampled 800-table chain
  200 paths with 0 intrusions; pin held across an LR→TB re-arrange; first-parse
  auto-arrange with 0 crossings reported;
• runtime archive structure and local-file startup smoke passing
  (procflow-v2.3.0.zip).

Upgrade and distribution

The runtime is self-contained. Use npm run package:runtime to produce the
versioned ZIP and verify .release/SHA256SUMS.txt before distributing it.
Layouts, workspaces, and layout files saved by earlier releases remain
readable.