v0.3.0

v0.3.0

Breaking for code and configurations written against 0.2.0:

• Requires Python 3.14 or later (#172). 0.2.0 remains the last release for Python 3.12 and 3.13.
• fetch_rba_series picks each month's value by observation date instead of file order, keeping the latest date in the month, and raises ValueError when one date carries two different values (#175). Identical duplicate dates are still accepted.
• Forecast evidence extraction refuses incomplete selected evidence and non-finite totals, and score_forecast requires a finite, positive weight total (#182). Missing values used to disappear during reduction, and two weights of 1e308 could erase nonzero errors. A refused portfolio candidate now fails its fold while the other folds continue.

Added:

• XeroReport.account_identities() and pyfpa.memory.reimport.compare_reimport() compare account identities, mappings and source metadata across Xero re-imports (#179). Account codes keep leading zeros, and the comparison reports additions, removals, renames, tracking changes, mapping changes and coverage gaps without changing amounts or company memory.

Fixed:

• write_onboarding_outputs refuses a linked decisions folder that leads out of the workspace, and replaces files through a sibling temporary file, so a failed overwrite no longer truncates an existing business profile (#173).
• A forecast snapshot whose assumptions cannot be written no longer creates an empty file or erases an existing snapshot during an explicit overwrite (#177).

v0.2.0

Breaking for code and configurations written against 0.1.2:

• Requires Python 3.12 or later, with pandas 3.0.5, PyYAML 6.0.3 and openpyxl 3.1.5 as the dependency floors (#159). 0.1.2 remains the last release for Python 3.11.
• The days-based working-capital model refuses a configuration that implies negative customer receipts, purchases or supplier payments, naming the month, the flow and the amount, and model-export writes no workbook for it. A DSO of 60 days with revenue tripling in a month used to report a smaller operating cash flow instead (#139). Use a collection profile or the replenishment basis below where the days model cannot follow the business.
• Forecast configuration rejects NaN and infinity (#135), a receipt shift rejects non-finite amounts and keeps cash unchanged when its source and destination month match (#136), and seasonality that overflows its total or growth that overflows raises ValueError instead of erasing revenue or producing infinite cash (#137).
• Removed public names nothing read: GstAssumptions.agent_lodgment (a YAML agent_lodgment key is now ignored), XeroReport.gst_inclusive, XeroReport.unmapped_tracking and the connector-validate --timeout flag (#134).

Added:

• review_forecast structural controls over a monthly forecast frame (#140), an opt-in income-year tax provision (#141), receivables collected by a stated profile (#142) with bad debts and later write-offs (#150), and an opt-in replenishment basis for inventory (#151).
• pyfpa.au.payroll_settlement settles payroll costs on their own dates (#143), with due-date rules for PAYG withholding (#144), the super guarantee under Payday Super and the quarterly regime (#145) and monthly payroll tax lodgers (#146), and PAYG instalments for quarterly payers (#149).
• Ridgeline cash decisions with a 12-cell receipt-delay and hiring grid (#166), and a Lumbridge close handoff that reconstructs the opening payable from its own fabricated bills, credits and settlements (#170).

Fixed:

• model-export refuses to write over its own configuration, including through links (#154); correction slugs cannot leave the corrections directory (#155); experiment saves no longer overwrite a competing save of the same name (#156); depreciation evidence refuses a non-boolean synthetic_input (#157); registry loaders report broken paths and non-mapping YAML instead of returning empty records (#158).
• The Lumbridge briefing's figures are tied to their sources by tests (#138), the learning workflows keep Xero API data out until the developer terms are checked (#148), and docs/ai-register-entry.md gives supplier information for a firm's AI register (#153).

v0.1.2

Breaking for scripts written against 0.1.1:

• reconcile-source and reconcile_account_table agree a target within half a cent
  by default (--abs-tolerance 0.005), and the relative --tolerance is off unless
  set. The old 1% default passed 3,000 cross-mapped between two revenue targets.
  Whole-dollar controls need --abs-tolerance 0.5.
• EntityConfig.tax_rate is required; the old 21% default was the US federal rate.
  start_month must be YYYY-MM, and date text given to payroll, GST, rate lookups and
  format_au_date must be ISO: day-first text such as 1/10/2026 is refused instead
  of being read as January.
• Config and 13-week cash files refuse unknown keys (EntityConfig and its parts,
  Cash13Config, WeeklyFlow), and opex amounts cannot be negative.
• read_xero_report refuses a report whose amount columns are not periods, such as a
  comparison or tracking export, unless tracking_comparison=True is set for a
  standard P&L with one tracking option per column. It also refuses a Total X row
  with no X heading above it and, in the flat layout, a row with more fields than
  the header, a filled row with no account, and unnamed or repeated columns.
  pyfpa.io.xero_au.from_xero is removed.
• Payroll tax and super guarantee lookups after 30 June 2027, the date those tables
  were verified to, raise until the tables are verified again.
• fetch_abs_series reads its key only from ABS_API_KEY and no longer takes an
  api_key argument. Its keys map to the ABS Indicator API's _H dataflows, a
  response in the wrong frequency is refused, and retail trade is history only: the
  ABS ceased it with the June 2025 period.
• Cross-client promotion is default-deny (below), seed_from_library requires
  company_root and seeded_at, and promote_challenger requires objective and
  approved_at.
• verify_workbook requires every expected column, a matching period header and zero
  check_* rows. recover_actuals raises for a line it cannot recover instead of
  skipping it, score_forecast skips zero-weight lines and refuses all-zero weights,
  and pyfpa.memory.diagnostics.validate_workspace is removed in favour of
  Workspace.validate.

Figures that change:

• A BAS due date on a weekend moves to the following Monday, and a leading partial
  quarter settles the months the series holds instead of being dropped.
• Workers compensation includes bonuses, as every state's wage definition does.
• fy_summary orders periods by date, and a divestiture keeps cogs consistent with
  gross profit.

New:

• 13-week cash scenarios: Scenario, apply_scenario, compare_scenarios and
  load_cash13_scenarios.
• pyfpa.au.depreciation reads accounting-depreciation evidence files, refuses one
  whose producer rejected it or whose digest no longer matches, and spreads a charge
  in whole cents with expense and asset purchases kept apart.
• verify_structure checks a workbook's layout contract, and model_to_excel writes
  check_* self-check rows.

Cross-client promotion:

• Cross-client promotion in pyfpa/portfolio is default-deny. promote_prior and
  promote_skill refuse unless a practitioner has recorded a PromotionApproval for
  that exact candidate's sha256 digest, listing the purpose, the contributing
  workspaces, the authorisation reference and a confidentiality review.
  record_promotion_approval is the only writer of an approval and refuses to
  overwrite one. The approval is an acknowledgement, not authentication and not legal
  proof of client consent.
• validate_prior accepts the candidate it validates and stamps its digest into
  ValidationResult; a result without the candidate's digest cannot support a
  promotion. screen_candidate reports ABN and TFN shaped numbers, email addresses,
  Australian phone numbers, BSB and account patterns, narrative dollar amounts and the
  contributing clients' business names, and any finding blocks promotion until the
  recorded review lists it.
• promote_skill copies only the files the approval lists in allowed_files. The gate
  counts distinct clients by normalised business-profile heading, ignoring case,
  punctuation, bracketed qualifiers and entity suffixes, and treats a workspace with no
  heading as establishing no client. Promoting across two workspaces that normalise to
  one name takes both workspace ids in the approval's aliases_acknowledged and review
  notes saying why they are separate.
• Shared library records now carry opaque workspace ids and the approval digest instead
  of client paths, with the path map in <library>/provenance/workspaces.yaml.
  seed_from_library takes the receiving company_root and seeded_at, records the
  seed in that workspace's .fpa/library-seeds.yaml and in the library's seed index,
  and the new withdraw_prior removes a prior and reports the workspaces it seeded
  without touching anything inside them.
• Approval files store contributing workspaces as opaque ids, so nothing in a shared
  library names a client directory. Building one still accepts paths. The library's seed
  index records the receiving workspace by id too, and withdraw_prior resolves ids back
  to paths through provenance/workspaces.yaml when it reports affected workspaces.
• record_promotion_approval writes with an exclusive create through the new
  pyfpa.io.loaders.create_yaml, so a second writer racing on the same digest is
  refused instead of replacing a recorded decision.
• validate_prior stamps each result it produces, and promote_prior refuses a result
  that stamp does not cover, which rejects a ValidationResult built by hand, edited on
  disk or carried over from another candidate. The stamp is a per-process structura...

v0.2.0

v0.2.0

Breaking for code and configurations written against 0.1.2:

• Requires Python 3.12 or later, with pandas 3.0.5, PyYAML 6.0.3 and openpyxl 3.1.5 as the dependency floors (#159). 0.1.2 remains the last release for Python 3.11.
• The days-based working-capital model refuses a configuration that implies negative customer receipts, purchases or supplier payments, naming the month, the flow and the amount, and model-export writes no workbook for it. A DSO of 60 days with revenue tripling in a month used to report a smaller operating cash flow instead (#139). Use a collection profile or the replenishment basis below where the days model cannot follow the business.
• Forecast configuration rejects NaN and infinity (#135), a receipt shift rejects non-finite amounts and keeps cash unchanged when its source and destination month match (#136), and seasonality that overflows its total or growth that overflows raises ValueError instead of erasing revenue or producing infinite cash (#137).
• Removed public names nothing read: GstAssumptions.agent_lodgment (a YAML agent_lodgment key is now ignored), XeroReport.gst_inclusive, XeroReport.unmapped_tracking and the connector-validate --timeout flag (#134).

Added:

• review_forecast structural controls over a monthly forecast frame (#140), an opt-in income-year tax provision (#141), receivables collected by a stated profile (#142) with bad debts and later write-offs (#150), and an opt-in replenishment basis for inventory (#151).
• pyfpa.au.payroll_settlement settles payroll costs on their own dates (#143), with due-date rules for PAYG withholding (#144), the super guarantee under Payday Super and the quarterly regime (#145) and monthly payroll tax lodgers (#146), and PAYG instalments for quarterly payers (#149).
• Ridgeline cash decisions with a 12-cell receipt-delay and hiring grid (#166), and a Lumbridge close handoff that reconstructs the opening payable from its own fabricated bills, credits and settlements (#170).

Fixed:

• model-export refuses to write over its own configuration, including through links (#154); correction slugs cannot leave the corrections directory (#155); experiment saves no longer overwrite a competing save of the same name (#156); depreciation evidence refuses a non-boolean synthetic_input (#157); registry loaders report broken paths and non-mapping YAML instead of returning empty records (#158).
• The Lumbridge briefing's figures are tied to their sources by tests (#138), the learning workflows keep Xero API data out until the developer terms are checked (#148), and docs/ai-register-entry.md gives supplier information for a firm's AI register (#153).

v0.1.2

Breaking for scripts written against 0.1.1:

• reconcile-source and reconcile_account_table agree a target within half a cent
  by default (--abs-tolerance 0.005), and the relative --tolerance is off unless
  set. The old 1% default passed 3,000 cross-mapped between two revenue targets.
  Whole-dollar controls need --abs-tolerance 0.5.
• EntityConfig.tax_rate is required; the old 21% default was the US federal rate.
  start_month must be YYYY-MM, and date text given to payroll, GST, rate lookups and
  format_au_date must be ISO: day-first text such as 1/10/2026 is refused instead
  of being read as January.
• Config and 13-week cash files refuse unknown keys (EntityConfig and its parts,
  Cash13Config, WeeklyFlow), and opex amounts cannot be negative.
• read_xero_report refuses a report whose amount columns are not periods, such as a
  comparison or tracking export, unless tracking_comparison=True is set for a
  standard P&L with one tracking option per column. It also refuses a Total X row
  with no X heading above it and, in the flat layout, a row with more fields than
  the header, a filled row with no account, and unnamed or repeated columns.
  pyfpa.io.xero_au.from_xero is removed.
• Payroll tax and super guarantee lookups after 30 June 2027, the date those tables
  were verified to, raise until the tables are verified again.
• fetch_abs_series reads its key only from ABS_API_KEY and no longer takes an
  api_key argument. Its keys map to the ABS Indicator API's _H dataflows, a
  response in the wrong frequency is refused, and retail trade is history only: the
  ABS ceased it with the June 2025 period.
• Cross-client promotion is default-deny (below), seed_from_library requires
  company_root and seeded_at, and promote_challenger requires objective and
  approved_at.
• verify_workbook requires every expected column, a matching period header and zero
  check_* rows. recover_actuals raises for a line it cannot recover instead of
  skipping it, score_forecast skips zero-weight lines and refuses all-zero weights,
  and pyfpa.memory.diagnostics.validate_workspace is removed in favour of
  Workspace.validate.

Figures that change:

• A BAS due date on a weekend moves to the following Monday, and a leading partial
  quarter settles the months the series holds instead of being dropped.
• Workers compensation includes bonuses, as every state's wage definition does.
• fy_summary orders periods by date, and a divestiture keeps cogs consistent with
  gross profit.

New:

• 13-week cash scenarios: Scenario, apply_scenario, compare_scenarios and
  load_cash13_scenarios.
• pyfpa.au.depreciation reads accounting-depreciation evidence files, refuses one
  whose producer rejected it or whose digest no longer matches, and spreads a charge
  in whole cents with expense and asset purchases kept apart.
• verify_structure checks a workbook's layout contract, and model_to_excel writes
  check_* self-check rows.

Cross-client promotion:

• Cross-client promotion in pyfpa/portfolio is default-deny. promote_prior and
  promote_skill refuse unless a practitioner has recorded a PromotionApproval for
  that exact candidate's sha256 digest, listing the purpose, the contributing
  workspaces, the authorisation reference and a confidentiality review.
  record_promotion_approval is the only writer of an approval and refuses to
  overwrite one. The approval is an acknowledgement, not authentication and not legal
  proof of client consent.
• validate_prior accepts the candidate it validates and stamps its digest into
  ValidationResult; a result without the candidate's digest cannot support a
  promotion. screen_candidate reports ABN and TFN shaped numbers, email addresses,
  Australian phone numbers, BSB and account patterns, narrative dollar amounts and the
  contributing clients' business names, and any finding blocks promotion until the
  recorded review lists it.
• promote_skill copies only the files the approval lists in allowed_files. The gate
  counts distinct clients by normalised business-profile heading, ignoring case,
  punctuation, bracketed qualifiers and entity suffixes, and treats a workspace with no
  heading as establishing no client. Promoting across two workspaces that normalise to
  one name takes both workspace ids in the approval's aliases_acknowledged and review
  notes saying why they are separate.
• Shared library records now carry opaque workspace ids and the approval digest instead
  of client paths, with the path map in <library>/provenance/workspaces.yaml.
  seed_from_library takes the receiving company_root and seeded_at, records the
  seed in that workspace's .fpa/library-seeds.yaml and in the library's seed index,
  and the new withdraw_prior removes a prior and reports the workspaces it seeded
  without touching anything inside them.
• Approval files store contributing workspaces as opaque ids, so nothing in a shared
  library names a client directory. Building one still accepts paths. The library's seed
  index records the receiving workspace by id too, and withdraw_prior resolves ids back
  to paths through provenance/workspaces.yaml when it reports affected workspaces.
• record_promotion_approval writes with an exclusive create through the new
  pyfpa.io.loaders.create_yaml, so a second writer racing on the same digest is
  refused instead of replacing a recorded decision.
• validate_prior stamps each result it produces, and promote_prior refuses a result
  that stamp does not cover, which rejects a ValidationResult built by hand, edited on
  disk or carried over from another candidate. The stamp is a per-process structural
  barrier, not a security control, so validate and promote in one session.
• validate_prior also refuses a candidate whose business type or value are not the ones
  its folds tested.
• promote_skill reads the skill tree once and writes those exact bytes, instead of
  copying the client's directory again after the screen, and refuses a symlink, junction
  or other reparse point inside the tree.
• seed_from_library digests each stored prior again from its driver, business type,
  value and contributing workspace ids, and refuses it unless that reproduces the
  recorded digest and its approval is on disk. A legacy entry with no digest, and an
  entry edited after promotion, are both refused.
• Loading an approval checks that the record's own candidate_digest is the one it was
  looked up by, so a record copied onto another candidate's filename is refused, and the
  findings stamp writes to the digest the gate verified.
• Documentation states that "nothing leaves your machine" is about network egress and
  does not by itself permit moving one client's information into another client's work.

v0.1.1

• Publishes the attested wheel and source distribution to PyPI as au-fpa-pack through trusted publishing.
• No functional change since v0.1.0.

v0.1.0

First release of the Australian FP&A pack for openfpa: 30 June financial years,
Xero AU mapping, GST and BAS cash timing, and payroll on-cost assumptions.

• openfpa, a JSON-emitting CLI for company workspace intake, source lineage,
  mappings,...

v0.1.2

v0.1.2

Breaking for scripts written against 0.1.1:

• reconcile-source and reconcile_account_table agree a target within half a cent
  by default (--abs-tolerance 0.005), and the relative --tolerance is off unless
  set. The old 1% default passed 3,000 cross-mapped between two revenue targets.
  Whole-dollar controls need --abs-tolerance 0.5.
• EntityConfig.tax_rate is required; the old 21% default was the US federal rate.
  start_month must be YYYY-MM, and date text given to payroll, GST, rate lookups and
  format_au_date must be ISO: day-first text such as 1/10/2026 is refused instead
  of being read as January.
• Config and 13-week cash files refuse unknown keys (EntityConfig and its parts,
  Cash13Config, WeeklyFlow), and opex amounts cannot be negative.
• read_xero_report refuses a report whose amount columns are not periods, such as a
  comparison or tracking export, unless tracking_comparison=True is set for a
  standard P&L with one tracking option per column. It also refuses a Total X row
  with no X heading above it and, in the flat layout, a row with more fields than
  the header, a filled row with no account, and unnamed or repeated columns.
  pyfpa.io.xero_au.from_xero is removed.
• Payroll tax and super guarantee lookups after 30 June 2027, the date those tables
  were verified to, raise until the tables are verified again.
• fetch_abs_series reads its key only from ABS_API_KEY and no longer takes an
  api_key argument. Its keys map to the ABS Indicator API's _H dataflows, a
  response in the wrong frequency is refused, and retail trade is history only: the
  ABS ceased it with the June 2025 period.
• Cross-client promotion is default-deny (below), seed_from_library requires
  company_root and seeded_at, and promote_challenger requires objective and
  approved_at.
• verify_workbook requires every expected column, a matching period header and zero
  check_* rows. recover_actuals raises for a line it cannot recover instead of
  skipping it, score_forecast skips zero-weight lines and refuses all-zero weights,
  and pyfpa.memory.diagnostics.validate_workspace is removed in favour of
  Workspace.validate.

Figures that change:

• A BAS due date on a weekend moves to the following Monday, and a leading partial
  quarter settles the months the series holds instead of being dropped.
• Workers compensation includes bonuses, as every state's wage definition does.
• fy_summary orders periods by date, and a divestiture keeps cogs consistent with
  gross profit.

New:

• 13-week cash scenarios: Scenario, apply_scenario, compare_scenarios and
  load_cash13_scenarios.
• pyfpa.au.depreciation reads accounting-depreciation evidence files, refuses one
  whose producer rejected it or whose digest no longer matches, and spreads a charge
  in whole cents with expense and asset purchases kept apart.
• verify_structure checks a workbook's layout contract, and model_to_excel writes
  check_* self-check rows.

Cross-client promotion:

• Cross-client promotion in pyfpa/portfolio is default-deny. promote_prior and
  promote_skill refuse unless a practitioner has recorded a PromotionApproval for
  that exact candidate's sha256 digest, listing the purpose, the contributing
  workspaces, the authorisation reference and a confidentiality review.
  record_promotion_approval is the only writer of an approval and refuses to
  overwrite one. The approval is an acknowledgement, not authentication and not legal
  proof of client consent.
• validate_prior accepts the candidate it validates and stamps its digest into
  ValidationResult; a result without the candidate's digest cannot support a
  promotion. screen_candidate reports ABN and TFN shaped numbers, email addresses,
  Australian phone numbers, BSB and account patterns, narrative dollar amounts and the
  contributing clients' business names, and any finding blocks promotion until the
  recorded review lists it.
• promote_skill copies only the files the approval lists in allowed_files. The gate
  counts distinct clients by normalised business-profile heading, ignoring case,
  punctuation, bracketed qualifiers and entity suffixes, and treats a workspace with no
  heading as establishing no client. Promoting across two workspaces that normalise to
  one name takes both workspace ids in the approval's aliases_acknowledged and review
  notes saying why they are separate.
• Shared library records now carry opaque workspace ids and the approval digest instead
  of client paths, with the path map in <library>/provenance/workspaces.yaml.
  seed_from_library takes the receiving company_root and seeded_at, records the
  seed in that workspace's .fpa/library-seeds.yaml and in the library's seed index,
  and the new withdraw_prior removes a prior and reports the workspaces it seeded
  without touching anything inside them.
• Approval files store contributing workspaces as opaque ids, so nothing in a shared
  library names a client directory. Building one still accepts paths. The library's seed
  index records the receiving workspace by id too, and withdraw_prior resolves ids back
  to paths through provenance/workspaces.yaml when it reports affected workspaces.
• record_promotion_approval writes with an exclusive create through the new
  pyfpa.io.loaders.create_yaml, so a second writer racing on the same digest is
  refused instead of replacing a recorded decision.
• validate_prior stamps each result it produces, and promote_prior refuses a result
  that stamp does not cover, which rejects a ValidationResult built by hand, edited on
  disk or carried over from another candidate. The stamp is a per-process structural
  barrier, not a security control, so validate and promote in one session.
• validate_prior also refuses a candidate whose business type or value are not the ones
  its folds tested.
• promote_skill reads the skill tree once and writes those exact bytes, instead of
  copying the client's directory again after the screen, and refuses a symlink, junction
  or other reparse point inside the tree.
• seed_from_library digests each stored prior again from its driver, business type,
  value and contributing workspace ids, and refuses it unless that reproduces the
  recorded digest and its approval is on disk. A legacy entry with no digest, and an
  entry edited after promotion, are both refused.
• Loading an approval checks that the record's own candidate_digest is the one it was
  looked up by, so a record copied onto another candidate's filename is refused, and the
  findings stamp writes to the digest the gate verified.
• Documentation states that "nothing leaves your machine" is about network egress and
  does not by itself permit moving one client's information into another client's work.

v0.1.1

• Publishes the attested wheel and source distribution to PyPI as au-fpa-pack through trusted publishing.
• No functional change since v0.1.0.

v0.1.0

First release of the Australian FP&A pack for openfpa: 30 June financial years,
Xero AU mapping, GST and BAS cash timing, and payroll on-cost assumptions.

• openfpa, a JSON-emitting CLI for company workspace intake, source lineage,
  mappings, reconciliation, connector scaffolding, corrections, scorecards,
  context packs and verified Excel export.
• 17 skills under skills/, plus the .claude-plugin/plugin.json manifest that
  declares the repository as the openfpa plugin.
• Five worked examples: Lumbridge Services, Harbour Light and Ridgeline Chair
  Co. from synthetic records, ARB Corporation from its FY2025 Appendix 4E, and
  Fox Factory Holding Corp. from public SEC filings.
• The distribution is au-fpa-pack and the import is pyfpa. Version 0.1.1
  is available on PyPI.
• Tested on Python 3.11, 3.12 and 3.13.

v0.1.1

v0.1.1

• Publishes the attested wheel and source distribution to PyPI as au-fpa-pack through trusted publishing.
• No functional change since v0.1.0.

v0.1.0

First release of the Australian FP&A pack for openfpa: 30 June financial years,
Xero AU mapping, GST and BAS cash timing, and payroll on-cost assumptions.

• openfpa, a JSON-emitting CLI for company workspace intake, source lineage,
  mappings, reconciliation, connector scaffolding, corrections, scorecards,
  context packs and verified Excel export.
• 17 skills under skills/, plus the .claude-plugin/plugin.json manifest that
  declares the repository as the openfpa plugin.
• Five worked examples: Lumbridge Services, Harbour Light and Ridgeline Chair
  Co. from synthetic records, ARB Corporation from its FY2025 Appendix 4E, and
  Fox Factory Holding Corp. from public SEC filings.
• Source-only. The distribution is au-fpa-pack and the import is pyfpa; it
  is not published to PyPI.
• Tested on Python 3.11, 3.12 and 3.13.

v0.1.0

v0.1.0

First release of the Australian FP&A pack for openfpa: 30 June financial years,
Xero AU mapping, GST and BAS cash timing, and payroll on-cost assumptions.

• openfpa, a JSON-emitting CLI for company workspace intake, source lineage,
  mappings, reconciliation, connector scaffolding, corrections, scorecards,
  context packs and verified Excel export.
• 17 skills under skills/, plus the .claude-plugin/plugin.json manifest that
  declares the repository as the openfpa plugin.
• Five worked examples: Lumbridge Services, Harbour Light and Ridgeline Chair
  Co. from synthetic records, ARB Corporation from its FY2025 Appendix 4E, and
  Fox Factory Holding Corp. from public SEC filings.
• Source-only. The distribution is au-fpa-pack and the import is pyfpa; it
  is not published to PyPI.
• Tested on Python 3.11, 3.12 and 3.13.