This is a comprehensive probate document from Edinburgh dated November 17, 1825, concerning the estate of Walter Irvine, Esquire, who died on January 7, 1824, at Luddington House, Egham, Surrey. He held properties in the Island of Tobago (West Indies), Luddington House in Surrey, and Denino in Fife, Scotland.

Scottish Estate Inventory (Total: £5,258 6s 2d)

I. Rents and Arrears Due (£4,569 13s 2½d):

• Robert Gray: £654 9s 6d
• James Quirie: £455 4s
• Thomas Wilson: £440 14s 4d
• Alcy Gosman: £982 10s 3d
• John Reckie: £763 5s 4d
• John Hortune: £342 15s 2d
• Balkaithly Park rents
• D Gosman feuduty: £8 18s 4d

II. Bond from John M. Harlaw of Kirkton: £245 5s 1d plus interest to January 1824 (Total: £262 15s 1½d)

III. Bond by Road Trustees (Anstruther-Crail to St. Andrews): £150 principal plus interest (Total: £162 7s 4½d)

IV. Bond by Lundenfoot Road Trustees: £50 with compounded interest from 1821-1824 (Total: £55 1s 2d)

V. Proceeds from sale of wood: £33 10s

VI. Balance due from Walter Dickson WS (Writer to the Signet): £174 16s 11d plus interest (Total: £174 18s 11d)

Will Provisions (dated October 18, 1823)

Primary Beneficiaries:

Wife - Catherine Irvine (née Gordon):

• Annual annuity of £1,500 for life (one-third paid from Scottish estates, two-thirds from daughters' inheritances)
• Lump sum of £1,000 (£500 within 3 months, £500 within 6 months of death)
• Lifetime use of all carriages, horses, household furniture, plate, linen, china, books, and consumable wines/liquors in Great Britain
• Right to occupy Luddington House for life, or receive rents if she chooses not to occupy (£400 annually if the property is sold)
• This provision replaces her marriage settlement rights and any dower claims

Three Legitimate Daughters (each receiving £35,000):

1. Elizabeth Douglas (married to William Robert Keith Douglas):

   • Subject to marriage settlement of November 23, 1821
   • Receives one-third value of all Walter Irvine's estates, calculated primarily from the Fife properties
   • Her portion held for her separate use, independent of her husband
   • She has full power to dispose of it by deed or will despite coverture (marriage)

2. Christina Charles and Catherine Grace (unmarried):

   • Each receives £35,000 legacy
   • Held in trust, with income for their separate use during their lives
   • Cannot be anticipated or attached by creditors or future husbands
   • If married, their surviving husbands receive life interest (only if marriage approved by trustees and if first husband)
   • Upon their deaths, portions pass to their children; if no children survive, portions revert to Elizabeth or are divided among sisters

Natural Daughter - Betsey (of London):

• Trust fund of £3,333 6s 8d in 3% consolidated annuities
• Receives income for life for her separate use
• After her death, capital passes to her children
• She and any husband can jointly appoint how funds are distributed to children
• Also receives £25 for mourning

Other Significant Bequests:

Ann Stewart (sister-in-law and widow of Stewart):

• £50 for mourning
• Original annuity of £100 increased to £200 per year for life
• Paid for her separate use, independent of any husband
• Funded by £6,668 13s 4d in 3% consolidated annuities

Brother's Estate:

• Forgives all debts owed by the estate of his late brother Christopher William Irvine

Nephews:

• Christopher Irvine (of Tobago): £4,000 payable one year after death with 4% interest thereafter
• Walter Irvine: £3,000 under same terms

Nieces (Clara and Marian Irvine - daughters of Christopher William):

• Each receives annuity of £100 per year for life
• Each funded by £3,333 6s 8d in 3% consolidated annuities

Jane Armstrong (formerly Jane Currie, wife of Francis Armstrong):

• £100 lump sum within 3 months for her separate use
• Annuity of £50 per year for life, paid for her separate use
• Funded by £1,666 13s 4d in 3% consolidated annuities

Mary (reputed daughter of a negro woman named Linda in Tobago):

• £500 sterling within 6 months of death

Other Named Beneficiaries:

• Alexander Gordon (brother-in-law and executor): £500 for his trouble executing the will (£100 within 6 months, £400 in annual installments)
• John Macfarlane, Esquire: £50
• Grace (wife of David Maitland), Daniel Gordon, Euphemia Gordon, and Elizabeth Gordon: £200 each plus £25 each for mourning
• Susannah Ann Hamilton (daughter of executor John Hamilton): £100 (added by codicil)

Continuing Annuities:

• Mrs. McIntyre: £5 annually during her life
• Two Miss Irvines: 3 guineas each annually during their lives
• All funded by appropriate 3% consolidated annuities

Trust Structure and Powers:

Executors and Trustees:

1. Catherine Irvine (wife)
2. William Robert Keith Douglas (son-in-law)
3. Alexander Gordon (brother-in-law, of Old Broad Street, London)
4. John Hamilton (of Kempshot Park, Hampshire)
5. Christopher Irvine (nephew, of Tobago)
6. Thomas Bruce (of George Muir, County of Fife)

Powers Granted to Trustees:

• Full discretion to manage and invest trust funds in British government stocks, real securities in England or Scotland
• Authority to purchase lands in fee simple or long leases (100+ years remaining)
• Power to convert land back to money through sales
• Can make leases up to 21 years
• May advance up to one-fourth of any child's portion for education, profession, military commission, or marriage
• Can use income from children's portions for their maintenance and education during minority
• William Robert Keith Douglas given principal management of West Indies estates
• Alexander Gordon entitled to usual solicitor's charges despite being a trustee
• Power to appoint new trustees upon death or resignation

Specific Provisions for Children's Trusts:

• Income held for separate use of each daughter, protected from husbands' debts
• Children of daughters inherit at age 21 (sons) or age 21/marriage (daughters)
• Compound interest applied to accumulate portions until vesting
• Complex survivorship provisions if daughters die without surviving children

West Indies Property:

John Leith Dispute:

• Walter Irvine had claims on property of late John Leith of Tobago
• Instructs executors to settle these claims quickly
• Prefers cash payment over taking possession of Old and New Grange estates in Tobago

Plantation Management:

• Douglas given control of West Indies estates, consignment of produce, and provision of supplies

Scottish Property Requirements:

Election Doctrine:

• All beneficiaries must execute proper Scottish deeds within 6 months of death (or majority for younger daughter)
• Those refusing forfeit benefits under the will to compensate disappointed parties
• Necessary to enable trustees to carry out the will's provisions in Scotland

Residuary Estate:

After all debts, legacies, and annuities paid, the residue divided equally among three legitimate daughters as tenants in common:

• Elizabeth Douglas
• Christina Charles
• Catherine Grace

Elizabeth's share for her separate use with power to dispose by deed or will despite coverture.

Probate Details:

Proved: January 28, 1824, in Prerogative Court of Canterbury

Initial Grant of Administration to:

• Catherine Irvine (relict/widow)
• William Robert Keith Douglas
• Alexander Gordon

Reserved Power for: John Hamilton, Christopher Irvine, and Thomas Bruce to apply later

Witnesses to Will:

• Colvell Wyatt Edgell (of Milton Place, Egham, Surrey)
• Robert Logan (Egham Lodge)
• William Adams (servant to Mr. Irvine)

Codicil: Dated same day (October 18, 1823), adding £100 bequest to Susannah Ann Hamilton

Scottish Confirmation: Presented November 17, 1825, before George Carshin Jr., Commissioner of Edinburgh, with deposition by W.R.K. Douglas confirming the inventory and will details.

The document demonstrates extensive wealth with significant properties in Scotland, England, and the West Indies, careful provision for family members including an illegitimate daughter, and sophisticated trust arrangements typical of wealthy estates in the early 19th century.

Document Date: 17 November 1825 (Edinburgh)
Testator: Walter Irvine, Esquire (late of Tobago, Luddington House, Surrey, and Denino, Fife)
Date of Death: 7 January 1824
Will Date: 18 October 1823 (with a codicil of the same date)
Probate: Granted 28 January 1824 by the Prerogative Court of Canterbury

1. Inventory of Personal Estate (Scotland)

Presented by executor William Robert Keith Douglas (WS) and signed by him and George Carfrae Jr. (Commissioner). The total value of the Scottish estate is £5,258 / 6 / 2 (£5,258.31).

The testator also owned property in England; letters of administration for that part were already taken out by the executors.

2. Deposition (25 October 1825)

Executor W. R. K. Douglas swore that:

• Walter Irvine died 7 January 1824.
• He, his widow Catherine Gordon or Irvine, Alexander Gordon, John Hamilton, Christopher Irvine, and Thomas Bruce entered upon possession as executors under the will proved 28 January 1824.
• He knew of no other settlement or writing relative to the estate except the will and the marriage contract of Walter Irvine and Catherine (date blank).
• The inventory together with the English property constitutes a full and complete inventory of the personal estate, so far as known to him.
• The Scottish estate is valued between £5,000 and £6,000.
• He does not require confirmation to be expedited.

3. Probate

• Proved 28 January 1824 before the Prerogative Court of Canterbury.
• Grant to three executors: Catherine Irvine (widow), William Robert Keith Douglas, and Alexander Gordon (all duly sworn).
• Power reserved to grant similar administration to the other named executors (John Hamilton, Christopher Irvine, Thomas Bruce) upon application.
• Certificate dated 12 November 1825 signed by W. R. K. Douglas confirming the probate of the will.

4. Last Will and Testament – Detailed Summary

4.1 General Directions

• Revokes all earlier wills and codicils.
• All just debts, funeral expenses, and legacies to be paid from personal estate; if insufficient, real estate is charged.
• All payments to be made with convenient speed unless otherwise directed.

4.2 Provision for Wife – Catherine Irvine

• Annuity: £1,500 per year (half‑yearly payments on 25 June and 25 December), charged on all estates.
  • Scottish estates settled for daughter Elizabeth bear one‑third of this annuity.
  • Each surviving daughter (or her heirs) must contribute a rateable third of the remaining two‑thirds.
• Residence: Catherine (and any unmarried daughters) may occupy Luddington House during her lifetime; if she ceases to occupy, the rents are paid to her. She may demise the house for up to three years after her death. If sold during her life, she receives £400 per year in lieu.
• Cash bequest: £1,000 (half within 3 months, half within 6 months).
• Household goods: Use for life of all carriages, carriage horses, furniture, plate, linen, china, books in Great Britain, plus wines and liquors consumed in her family.
• This provision is in lieu of her marriage settlement and bars all dower/thirds rights in England and Scotland.

4.3 Remission of Debt

• Remits all sums due from the estate of his late brother Christopher William Irvine for the benefit of those liable.

4.4 Bequests to Individuals

4.5 Trust for Natural Daughter Betsy

• Capital: £3,333 / 6 / 8 in 3% Consolidated annuities.
• Income: Paid to Betsy for life, for her sole and separate use, free from any husband’s control, without power to anticipate.
• Remainder: After Betsy’s death, the fund is held in trust for her children/issue as she (and any husband) may jointly appoint; failing appointment, for her children equally (sons at 21, daughters at 21 or marriage). If all such children fail under age without issue, the fund falls into the residue.

4.6 Main Trusts for Daughters Christina Charles and Catherine Grace

• Legacy: Each receives a legacy of £35,000, to be answered by purchasing 3% Consolidated annuities at the highest market price on the day of death (or last open day). The annuities carry the first dividend after death.
• Trust: During each daughter’s life, the income is for her sole and separate use, not subject to anticipation or her husband’s control.
• Surviving husband: If a husband survives, he is entitled to the income for life only if he was the first husband and married with the written consent of the trustees (or at least three of them).
• Remainder: After death of daughter and any surviving husband, the capital is held for her children/issue as she appoints by deed or will; in default, for her children equally (sons at 21, daughters at 21 or marriage). Further provisions deal with failure of issue (gift over to the other daughter or to the married daughter Elizabeth Douglas).
• Powers: Trustees may apply income for maintenance/education of minor grandchildren; may advance up to one‑quarter of a child’s share for profession, army commission, university, etc. (treated as a deduction). Undisposed income is to be compounded and added to principal.

4.7 Investment and Management Powers of Trustees

• May invest in British stocks/funds or real securities in England, Scotland, or Wales.
• With consent of the life tenant (or after death, at trustees’ discretion) may vary securities.
• Upon request of the life tenant, may purchase land (fee simple or term with ≥100 years unexpired); may later sell and reconvert to money; may lease lands for up to 21 years; apply rents as if they were interest.
• The fund may be held on the same trusts as personal estate.

4.8 Devise of Real Estate

• All freehold, copyhold, leasehold manors, plantations, negroes, slaves, lands, etc., in England, Scotland, West Indies, and elsewhere, plus estates held as mortgagee/trustee, are devised to the trustees (Catherine Irvine, W. R. K. Douglas, Alexander Gordon, John Hamilton, Christopher Irvine, Thomas Bruce) according to their nature, on the trusts declared in the will.

4.9 Executors and Guardians

• The same six persons are appointed executors in trust and guardians of the youngest daughter during her minority.

4.10 Election Regarding Scottish Estates

• All three daughters (and other beneficiaries) must, within six calendar months after death (or for the youngest daughter, within six months after coming of age), execute a proper Scotch form deed of disposition of all Scottish real/heritable estates to enable the trustees to carry out the will.
• Failure to do so binds them, under the doctrine of Election (as observed by English Courts of Equity), to compensate those disappointed by their neglect.

4.11 Covenant for Daughter Elizabeth’s Marriage Settlement

• The testator had covenanted (in a settlement dated ~23 November 1821) to settle one‑third of all his real and personal estate (worldwide) upon Elizabeth and her family.
• That covenant is subject to his lifetime power to dispose of property by act or will, and to charge it with annuities/legacies.
• The trustees (Catherine, Alexander, John, Christopher, Thomas) have sole discretion to arrange which part of the West Indies estates shall be appropriated to make up that third. Their certificate is binding.
• The third is computed after deducting all monetary legacies and annuity funds (except those created solely for life annuities). This third is the primary fund to pay one‑third of the wife’s annuity, exonerating Elizabeth’s share.
• Trustees are empowered to cause valuations and to complete the settlement accordingly.

4.12 Residuary Estate

• After payment of debts, legacies, and fulfilling the marriage settlement, all remaining real and personal estate (including funds appropriated for annuities) is held in trust for the three daughters (Elizabeth Douglas, Christina Charles, Catherine Grace) equally as tenants in common (not joint tenants).
• Elizabeth’s share is for her sole and separate use; she may dispose of it by deed or will as if a feme sole, regardless of coverture.

4.13 Receipts and Discharges

• Receipts in writing by any trustee (or their survivors, heirs, agents) for money payable under the will shall effectually discharge the payer from seeing to application or being accountable for misapplication.

4.14 Powers of Executors

• Executors may pay debts on any evidence they think proper; accept security; compromise debts owed to the estate; allow time for payment.
• They are directed to accept settlement of claims on John Leith’s Tobago property in preference to taking the Old and New Grange estates in Tobago.

4.15 Appointment of New Trustees

• If any trustee dies, is discharged, or refuses to act, the surviving/acting trustees may appoint new trustees by deed under seal, attested by two witnesses.
• The trust estates shall be conveyed to the new trustees, who shall have the same powers as original trustees.

4.16 Indemnity of Trustees

• Trustees may deduct and retain from trust monies all costs, charges, and expenses incurred in execution of the trusts (approved by the majority of acting trustees).
• Alexander Gordon, although a trustee, may charge the usual solicitor’s fees for business transacted by him.
• Trustees are only chargeable for their own receipts and wilful defaults; not liable for loss without wilful default.

4.17 Management Powers

• W. R. K. Douglas is entrusted with principal direction and management of the West Indies estates (consignment of produce, supply of plantations).
• Trustees may abate rents, make leases up to 21 years of untenanted or soon‑expiring trust property.

4.18 Execution

• Signed and sealed by Walter Irvine on 18 October 1823; witnessed by Colgell Wyatt Edgell, Robert Logan, and William Adams (servant to Mr. Irvine).

5. Codicil (18 October 1823)

• Adds a bequest of £100 to Susanna Ann Hamilton, daughter of John Hamilton Esquire.

6. Affidavits and Probate Details

Handwriting Affidavits (for the codicil)

• William Matthews and Joseph Pennington (both of Old Broad Street, London) swore they knew Walter Irvine’s handwriting and believed the signature on the codicil to be his.
• Sworn before John Daubeny, Surrogate, on 21 and 22 January 1824.

Attesting Witness Affidavit

• William Adams (servant) swore he was present at the execution of the will; described the signing, sealing, publishing, and witnessing; confirmed the testator was of sound mind. Sworn 21 January 1824.

Grant of Probate

• Granted to Catherine Irvine, W. R. K. Douglas, and Alexander Gordon on 28 January 1824.
• Power reserved to the other executors (John Hamilton, Christopher Irvine, Thomas Bruce) to apply later.

7. Key Figures at a Glance

• Total Scottish personal estate: £5,258 / 6 / 2
• Wife’s annuity: £1,500/year
• Two younger daughters’ legacies: £35,000 each (to be invested in 3% consols)
• Natural daughter Betsy’s fund: £3,333 / 6 / 8 3% consols
• Nephews (Christopher & Walter Irvine): £3,000 each
• Nieces (Clara & Marian): £100/year each (funded by £3,333 / 6 / 8 each)
• Jane Armstrong: £100 + £50/year annuity
• Mary (daughter of Linda): £500

This document is a comprehensive and intricate testament disposing of a substantial trans‑Atlantic estate, providing generously for the testator’s wife, daughters (both legitimate and natural), other relatives, and servants, while establishing elaborate trusts to preserve wealth across generations.

This document is a collection of probate records concerning the estate of Walter Irvine, an Esquire who held property in the Island of Tobago (West Indies), Luddington House in Surrey, and Denino in Fife, Scotland. He died at Luddington House on 7 January 1824. The documents date from between late 1823 and late 1825 and comprise an inventory of his Scottish personal estate, a sworn deposition, his last will and testament (dated 18 October 1823) with a codicil, supporting affidavits, and the formal grant of probate from the Prerogative Court of Canterbury.

I. Inventory of the Personal Estate (Scotland)

The inventory was presented in Edinburgh on 17 November 1825 by Mr. Walter Dickson, Writer to the Signet, and made out by William Robert Keith Douglas, Esquire, one of the Lords of the Admiralty and an executor under the will (which was proved in the Prerogative Court of Canterbury in 1824, under the Act 48 Geo III cap 140).

I. Rents and Arrears of Rent due to the deceased

II. Balance of debt due by John Hay Harlaw, Esquire of Kirkton

• Principal (at 4 Aug 1822): £245. 5. 1
• Interest to 7 Jan 1824 at 5%: £17. 10. 0½
• Total: £262. 15. 1½

III. Sum in bond by the Trustees of the Anstruther & Crail to St. Andrews road

• Principal: £150. 0. 0
• Interest (Whitsun 1822 to 7 Jan 1824, 5%): £12. 7. 4½
• Total: £162. 7. 4½

IV. Sum lent to the Trustees of the Lauderfoot road

• Principal: £50. 0. 0
• Interest at varying rates (5%, 4½%, 4% over successive years): £4. 17. 5
• Total: £55. 1. 2

V. Proceeds of roupes of wood due at death: £33. 10. 0

VI. Balance due by Walter Dickson WS on account current (31 Dec 1823)

• Principal: £174. 16. 11
• Interest at 3%: £0. 2. 0
• Total: £174. 18. 11

Total value of the deceased's estate in Scotland: £5258. 6. 2

The deceased also died possessed of property in England, for which Letters of Administration had already been taken out by the executors. The inventory was signed by W. R. K. Douglas and George Carfrae Jr.

II. Deposition (Edinburgh, 25 October 1825)

Before George Carfrae Junior, Commissioner appointed by the Commissaries of Edinburgh, William Robert Keith Douglas swore that:

• Walter Irvine died on 7 January 1824.
• Douglas, acting with Mrs. Catherine Gordon or Irvine (the widow), Alexander Gordon (Old Broad Street, London), John Hamilton (Kempshott Park, Hants), and Christopher Irvine (of Tobago), entered into possession and management of the estate as executors under the will dated 18 October 1823.
• The will was proved in the Prerogative Court of Canterbury on 28 January 1824 (the probate was then in London but would be transmitted before the inventory was recorded).
• He knew of no other settlement or writing disposing of the estate other than the will and an (undated) contract of marriage between the deceased and Catherine Gordon.
• The inventory (together with the English effects) was a full and complete inventory of the deceased's personal estate.
• The Scottish estate was valued at over £5,000 but under £6,000 sterling.
• The deponent did not require Confirmation to be expended for the specified debts/effects.

III. The Last Will and Testament (18 October 1823)

The will revokes all prior wills, codicils, and testamentary dispositions. It directs all just debts, funeral expenses, probate costs, and legacies to be paid promptly out of the personal estate, with the real estate charged for any deficiency.

Provision for the Wife — Catherine Irvine

• An annuity of £1,500 per annum for life, paid half-yearly on 25 June and 25 December.
• Charged on all estates; the Scottish estates settled on daughter Elizabeth bear one-third of it, and each of the three surviving daughters bears a rateable third of the remaining two-thirds.
• The wife and unmarried daughters may occupy Luddington House for life; if she vacates, rents go to her to fund another residence, and trustees may lease the property for terms ending within three years after her death. If sold in her lifetime, she receives £400 a year in lieu.
• A legacy of £1,000 (paid in two moieties within three and six months).
• Use of carriages, carriage horses, household furniture, plate, linen, china, books, and part of the wines/liquors for life.
• This provision is in lieu of her marriage settlement and bars all dower/thirds rights in both England and Scotland.

Remission of Debt

• All sums due from the estate of his late brother Christopher William Irvine are remitted for the benefit of those liable.

Bequest to Ann Stewart (sister-in-law)

• £50 for mourning.
• The existing bond annuity of £100 p.a. is ratified, and an additional £100 p.a. (already being paid) is confirmed for life — a total of £200 p.a.
• A fund of £6,668 13s 4d of 3% Consolidated annuities is appropriated as the only fund to answer this annuity.

Other Bequests

• John Macfarlane, Esquire: £50.
• Sister-in-law Grace (wife of David Maitland), Daniel Gordon, Euphemia Gordon, and Elizabeth Gordon: £200 each plus £25 each for mourning.
• Alexander Gordon (brother-in-law): £500 for trouble in executing the trusts (£100 within six months, the rest in annual instalments of £100).

Trust for Natural Daughter "Betsy"

• A trust of £3,333 6s 8d of 3% Consolidated annuities for his reputed daughter Betsy of London, for life, for her sole and separate use, free from any husband's control, without power to anticipate.
• After her death, the fund goes to her children/issue as she and any husband may jointly appoint (by deed or will); in default, equally among her lawful children at 21 (sons) or 21/marriage with consent (daughters), as tenants in common, with survivorship.
• If she has no children, the fund falls into residue.
• £25 extra for mourning for Betsy.

Bequest to Jane Armstrong (formerly Jane Currie)

• A lump sum of £100 (paid within three months) for her sole and separate use, free from her husband Francis Armstrong.
• An annuity of £50 p.a. for life, half-yearly on 5 July and 5 January, for her separate use, backed by a fund of £1,666 13s 4d of 3% annuities as the only fund.

Bequests to Nieces Clara and Marian Irvine

• Each of the two daughters of his late brother Christopher William Irvine receives an annuity of £100 p.a. for life (half-yearly on 5 July and 5 January), each backed by a fund of £3,333 6s 8d of 3% annuities as the only fund.
• The wife retains a general lien on all residuary estates for her annuity notwithstanding the appropriations.

Bequest to Mary (reputed daughter of enslaved woman Linda)

• £500 (paid within six months) to Mary, the reputed daughter of the negro woman Linda living in Tobago.

Annuities to Others

• The annual £5 to Mrs. McIntyre and the annual 3 guineas to the two Miss Irvines are continued for life, free of stamp duties, each backed by appropriated 3% annuities as the only fund.

Bequests to Nephews

• Christopher Irvine and Walter Irvine each receive £3,000 (paid one year after death, with 4% interest from that point until payment).

Main Legacies for Daughters Christina and Catherine

• Two separate legacies, each of £35,000, valued by 3% Consolidated annuities at the highest market price on the day of death (or the last open day if shut), carrying dividends from the first dividend day after death.
• One legacy is held in trust for daughter Christina Charles, the other for Catherine Grace, on protective trusts: income for each daughter's sole and separate use during life, without power to anticipate, free from any husband's control.
• A surviving husband may take income for life (provided he married with the written consent of the trustees).
• After each daughter and her surviving husband die, the legacy goes to her children/issue as she appoints by deed or will; in default, equally among children as tenants in common (sons at 21, daughters at 21 or marriage), with survivorship provisions.
• On failure of issue, portions are divided between the other daughters' shares, ultimately accumulating toward daughter Elizabeth's settled portion.

Powers of Trustees (Maintenance and Advancement)

• Trustees may apply income for minor children's maintenance, education, schooling, clothing, or advancement.
• They may advance up to one-fourth of a child's vested or expectant share for placing a male into a profession, purchasing a military commission, university education, Inns of Court, or a child's marriage/advancement.
• Advances are deducted from the child's portion; capitalized income accrues with compound interest and follows the same trusts.

Investment Powers

• With the life-tenant's consent (and after their death, on trustees' own authority), trustees may invest in British stocks/funds or real securities in England, Scotland, or Wales.
• They may vary/change securities, and (on request of life tenants) purchase or sell land held in fee simple or with at least 100 years unexpired, lease land (up to 21 years), and hold it on the same trusts as personal estate.

Devise of Real Estate

• The will devises all freehold, copyhold, and leasehold manors, messuages, farms, plantations, negroes, and slaves, lands and hereditaments in England, Scotland, and the West Indies (and estates held as mortgagee/trustee), and all goods and personal estate, to the six trustees (Catherine Irvine, W. R. K. Douglas, Alexander Gordon, John Hamilton, Christopher Irvine, and Thomas Bruce) on the stated trusts.

Appointment of Executors and Guardians

• The six named individuals are appointed executors in trust and guardians of the youngest daughter during her minority.

Election Requirement for Scottish Estates

• The three daughters and all beneficiaries must, within six months (or, for the youngest daughter, within six months of coming of age), execute a Deed in Scotch form disposing of the Scottish real/heritable estates so trustees can carry the will into effect.
• Anyone refusing is bound, under the doctrine of Election, to give compensation/equivalence out of their benefits.

Covenant Regarding Daughter Elizabeth's Marriage Settlement

• The will directs performance of the covenant in the marriage settlement of 23 November 1821 (between W. R. K. Douglas, Walter Irvine and Elizabeth, and Robert Bruce of Kennet and John Charles Herries) for settling on Elizabeth one-third of all real and personal property in Great Britain, Ireland, the West Indies, and America.
• This third is to be made up primarily from the Fife estate (regardless of whether it exceeds one-third in value), supplemented if necessary from West Indian plantations/negroes/slaves.
• The trustees (or any three) have sole discretion over the valuation and appropriation of West Indian estates; their certificate is binding and conclusive.
• The settled third is the primary fund to answer one-third of the annuities, in exoneration of Elizabeth, and the trustees are to complete the settlement following valuation.

Residuary Estate Distribution

• Subject to debts, legacies, and regulations, all real and personal estate (including funds appropriated for annuities) is held in trust for the three daughters — Elizabeth Douglas, Christina Charles, and Catherine Grace — or those surviving, equally as tenants in common (not joint tenants).
• Elizabeth's share is for her sole and separate use, with full power to dispose of it by deed or will as if a feme sole.
• Reversions in settled Scottish/West Indian estates and the two £35,000 portions pass to the two unmarried daughters as tenants in common with survivorship on failure of issue.
• No deduction is made from Elizabeth's one-third in respect of her residuary share.

Receipts, Powers of Executors, New Trustees, Indemnity, Management

• Receipts: Trustees' written receipts are sufficient discharges.
• Executors' powers: pay disputed debts as they think proper, accept securities, compromise debts, allow time for payment; and if his claims on the estate of John Leith of Tobago are settled, executors should prefer cash payment over taking the Old and New Grange estates in Tobago.
• New trustees: may be appointed by deed on death/refusal to act, with full powers.
• Indemnity: trustees may deduct costs from trust funds; Alexander Gordon (as a solicitor) may make usual solicitor's charges; trustees are liable only for their own acts/wilful defaults, not for losses without wilful default.
• Management: trustees may entrust W. R. K. Douglas with the principal direction of the West Indian estates, consignment of produce, and supplies; may abate rents and grant leases up to 21 years.

Execution and Attestation

Executed on 18 October 1823 at Luddington House, signed and sealed by Walter Irvine across 19 sheets, in the presence of witnesses:

• Colgell Wyatt Edgell (Milton Place, Egham, Surrey)
• Robert Logan (Egham Lodge)
• William Adams (servant to Mr. Irvine)

IV. Codicil

Dated 18 October 1823, adds a bequest of £100 to Susanna Ann Hamilton, daughter of John Hamilton, Esquire.

V. Affidavits and Grant of Probate

Affidavit (Codicil)

• William Matthews (Old Broad Street, London) and Joseph Pennington swore they knew Walter Irvine's handwriting and believed the signature "Walter Irvine" on the codicil (annexed, marked "A") to be genuine. Pennington was sworn on 21 January 1824, and Matthews on 22 January 1824, before John Daubeny, Surrogate, and John I. Glennie, Notary Public.

Affidavit (William Adams, subscribing witness)

• William Adams swore he witnessed the testator sign the will on the nineteenth sheet opposite a wax seal, that the testator sealed, published, and declared it as his last will, and that he (Adams) was of sound and disposing mind, memory, and understanding at the time. Sworn 21 January 1824.

Grant of Probate (28 January 1824)

• Charles, Archbishop of Canterbury, grants probate in the Prerogative Court of Canterbury before John Daubeny, LL.D., Surrogate of Sir John Nicholl, Knight, Master Keeper or Commissary.
• The deceased is described as formerly of the Island of Tobago, of Wimpole Street, Middlesex, and late of Luddington House, Egham, Surrey.
• Administration was granted to Catherine Irvine (the relict), William Robert Keith Douglas, and Alexander Gordon, three named executors, who swore to well and faithfully administer and exhibit a true inventory by 31 July next.
• Power reserved to later grant administration to the other executors (John Hamilton, Christopher Irvine the nephew, and Thomas Bruce) when they apply.
• Signed by registrars Nathaniel Goschius, N. C. Cresswell, and George Firmer (Deputy Registers).

VI. Certificate

London, 12 November 1825 — William Robert Keith Douglas certifies that this is the probate of the will of the late Walter Irvine referred to in his 25 October 1825 deposition on the inventory.

Key Points at a Glance

• Testator: Walter Irvine, Esquire — properties in Tobago (West Indies), Luddington House (Surrey), and Denino (Fife); also Wimpole Street, Middlesex. Died 7 January 1824.
• Scottish personal estate value: £5,258 6s 2d (over £5,000, under £6,000); additional English assets subject to Letters of Administration.
• Executors/Trustees (6): Catherine Irvine (wife), William Robert Keith Douglas, Alexander Gordon, John Hamilton, Christopher Irvine (nephew), Thomas Bruce.
• Major beneficiaries: wife (annuity of £1,500 p.a.); three daughters — Elizabeth Douglas (marriage settlement third), Christina Charles and Catherine Grace (each £35,000); natural daughter Betsy (life interest trust); two nephews (each £3,000); plus numerous smaller legacies and life annuities to family, relations, and dependants (including a sum to Mary, reputed daughter of the enslaved woman Linda).
• Notable features: protective trusts to shield women's income from husbands; appropriation funds (3% Consols) dedicated to each annuity; the "doctrine of Election" for Scottish estates; power to purchase military commissions; and a slavery-era estate composed of plantations, negroes, and slaves in the West Indies.

Step 3.7 Implementation Analysis

1. Deviations from Plan/Proposal

1.1 Windows 1-byte and 2-byte operations use hash table fallback

Plan: "Windows: always bypass (all sizes 1–8 bytes). The Windows backend never uses the hash table."

Actual: atomic_wait_windows.c.ipp only defines HAVE_WAIT_ON_ADDRESS_32 and HAVE_WAIT_ON_ADDRESS_64. There are no wait_on_address8, wait_on_address16, wake_by_address8, or wake_by_address16 definitions. As a result, atomic_wait_1, atomic_wait_2, atomic_notify_one_1, atomic_notify_one_2, atomic_notify_all_1, and atomic_notify_all_2 on Windows fall through to atomic_wait_generic / atomic_notify_generic and use the hash table proxy.

Files: include/wg14_atomic_waits/detail/impl/atomic_wait_windows.c.ipp:42-74

1.2 macOS atomic_notify_all return value is platform-fixed to INT_MAX - 1

Plan: "Returns a positive number (possibly 1 + N where the extra info depends on platform)."

Actual: Both wake_by_address32 and wake_by_address64 on macOS return INT_MAX - 1 when max_threads_to_wake != 1. __ulock_wake does not return a count of woken threads, so the implementation cannot know how many were woken. Returning a fixed large value is not consistent with the plan's "possibly one plus the number of threads woken" semantics.

Files: include/wg14_atomic_waits/detail/impl/atomic_wait_macos.c.ipp:106, include/wg14_atomic_waits/detail/impl/atomic_wait_macos.c.ipp:123

1.3 Windows atomic_notify_all calls WakeByAddressSingle instead of WakeByAddressAll

Plan: "On Windows, call WakeByAddressSingle / WakeByAddressAll directly without any wake_waiters helper."

Actual: wake_by_address32 and wake_by_address64 always call WakeByAddressSingle, regardless of max_threads_to_wake. WakeByAddressAll is never called.

Files: include/wg14_atomic_waits/detail/impl/atomic_wait_windows.c.ipp:100, include/wg14_atomic_waits/detail/impl/atomic_wait_windows.c.ipp:117

1.4 atomic_wait_common.ipp.ipp lacks warning suppression for volatile-to-void* cast

Plan: "Test under -Wextra to suppress qualifier-discard warnings."

Actual: The cast void *const key = (void *) object; on line 173 discards both volatile and const qualifiers. With -Wextra, this produces a warning. No #pragma or explicit cast-to-uintptr_t-then-to-void* is used to suppress it.

Files: include/wg14_atomic_waits/detail/impl/atomic_wait_common.ipp.ipp:173

1.5 atomic_wait_expected_32 timeout error handling is incorrect on Linux

Plan: "Return negative on error."

Actual: wait_on_address32 on Linux always returns -1 on any failure, losing the distinction between ETIMEDOUT and other errors. atomic_wait_expected_32 then returns -1 with errno = 1 (because -ret2 = -(-1) = 1) for any error, including timeout. The code compares ret2 != ETIME && ret2 != ETIMEDOUT, but ret2 is always -1, so the check never matches the positive ETIMEDOUT value.

Files: include/wg14_atomic_waits/detail/impl/atomic_wait_linux.c.ipp:50, include/wg14_atomic_waits/detail/impl/atomic_wait_common.ipp.ipp:653-659

2. Bugs

2.1 Hash table proxy state machine is permanently stuck after first notify

Severity: Critical

Description: The proxy waiter uses proxy->atomic as a binary semaphore:

• Waiter parks on proxy->atomic == 0
• Notifier sets proxy->atomic = 1 and calls wake

After any notify, proxy->atomic is permanently 1. It is never reset to 0 for an existing proxy. The proxy is only initialized to 0 when first created in hash_table_find_or_create.

When a subsequent waiter arrives:

1. It acquires the hash table lock and finds the existing proxy (use_count incremented).
2. It unlocks and calls the platform wait primitive on proxy->atomic with expected value 0.
3. Since proxy->atomic == 1 (not 0), the wait returns immediately (EAGAIN on Linux/futex, non-match on Windows/macOS).
4. The waiter reloads *object. If it still equals expected (which it will if no one notified again), the loop repeats.
5. Step 3 repeats forever: proxy->atomic is still 1, so the wait returns immediately again.

Result: Infinite busy-loop on Windows/macOS; atomic_wait_generic returns -1 on Linux.

Root cause: There is no mechanism to reset proxy->atomic back to 0 after all waiters have consumed a notify.

Files: include/wg14_atomic_waits/detail/impl/atomic_wait_common.ipp.ipp:38-45, include/wg14_atomic_waits/detail/impl/atomic_wait_common.ipp.ipp:220-222

2.2 atomic_wait_generic returns -1 on proxy spurious wake instead of looping

Severity: High

Description: In atomic_wait_generic, after calling the proxy wait primitive:

const int ret2 = WG14_ATOMIC_WAITS_HASH_TABLE_ITEM_PROXY_TYPE_WAIT(item, ...);
if(ret2 < 0)
{
  return ret2;
}

If the proxy wait returns a negative value (e.g., -EAGAIN because proxy->atomic != 0), the function returns -1 immediately. This is incorrect: a spurious wake or a mismatched proxy state should trigger the re-compare-and-re-park loop, not an error return.

Files: include/wg14_atomic_waits/detail/impl/atomic_wait_common.ipp.ipp:411-418

2.3 atomic_wait_expected_32 returns -1 on proxy error instead of looping

Severity: High

Description: On non-native backends (or when using the hash table fallback), atomic_wait_expected_32 delegates to atomic_wait_generic. If atomic_wait_generic returns -1 due to the proxy state bug (section 2.1), atomic_wait_expected_32 propagates the -1 to the caller. Per the proposal, the only negative return should be on actual system errors, not on spurious proxy state mismatches.

Files: include/wg14_atomic_waits/detail/impl/atomic_wait_common.ipp.ipp:663-664

2.4 wait_on_address32 (Linux) discards errno, preventing timeout detection

Severity: Medium

Description: The Linux wait_on_address32 wrapper:

int ret = (int) syscall(SYS_futex, ...);
if(ret == 0 || errno == EAGAIN || errno == EINTR)
{
    errno = save_errno;
    return 0;
}
errno = save_errno;
return -1;

It always returns -1 on failure, regardless of whether the failure was ETIMEDOUT, EINTR, or something else. Callers cannot distinguish timeout from other errors.

Files: include/wg14_atomic_waits/detail/impl/atomic_wait_linux.c.ipp:43-51

2.5 atomic_notify_generic creates a proxy even when no waiters exist

Severity: Low

Description: atomic_notify_generic calls hash_table_find_or_create(table, object, false). If no proxy exists, this creates one with use_count = 0, stores 1 to proxy->atomic, and calls wake. The proxy is left in the hash table with use_count = 0 and proxy->atomic = 1. When the first waiter later arrives, it inherits the stuck proxy->atomic = 1 state, triggering the infinite loop described in section 2.1.

Files: include/wg14_atomic_waits/detail/impl/atomic_wait_common.ipp.ipp:443-446

3. Race Conditions (Lost Wakes)

3.1 Lost-wake race in hash table proxy between notify and waiter park

Severity: Critical

Timeline:

1. Thread A calls atomic_wait, loads *object == expected, enters the hash table path.
2. Thread A is preempted before calling hash_table_find_or_create.
3. Thread B calls atomic_notify_one (or atomic_notify_all). It acquires the hash table lock, looks up the proxy. If none exists, one is created with use_count = 0. It stores 1 to proxy->atomic and calls wake_by_address32. No one is waiting, so the wake has no effect.
4. Thread B releases the lock.
5. Thread A resumes, acquires the lock, finds the proxy (or creates one), increments use_count to 1, releases the lock.
6. Thread A calls the platform wait on proxy->atomic with expected 0.
7. Since proxy->atomic == 1 (set by Thread B in step 3), the wait returns immediately.
8. Thread A reloads *object. If still == expected (which it is, because Thread B may not have changed the value), Thread A loops.
9. Thread A calls wait again. proxy->atomic is still 1. Wait returns immediately again.
10. Result: Thread A loops forever (Windows/macOS) or returns -1 error (Linux).

Root cause: The notify path sets proxy->atomic = 1 as a "wake token", but there is no handshake to ensure the waiter has observed the token before it is reused. The proxy has no generation counter or sequence number.

Files: include/wg14_atomic_waits/detail/impl/atomic_wait_common.ipp.ipp:220-222, include/wg14_atomic_waits/detail/impl/atomic_wait_common.ipp.ipp:41-45

3.2 Lost-wake race in native paths between load-check and park

Severity: High

Timeline (native Linux/macOS/FreeBSD/Windows):

1. Thread A calls atomic_wait, loads *object == expected.
2. Thread A is preempted.
3. Thread B calls atomic_notify_one without changing *object (e.g., notify_all in a scenario where the value is already the desired state).
4. Thread B calls the platform wake primitive (FUTEX_WAKE, __ulock_wake, etc.) on object. No thread is currently parked, so the wake has no effect.
5. Thread A resumes and calls the platform wait primitive (FUTEX_WAIT, __ulock_wait, etc.) on object with expected value.
6. The kernel parks Thread A because *object == expected.
7. No further notify is forthcoming. Thread A blocks forever.

Why the while-loop doesn't help: The while-loop (while(load == expected) { wait(); }) only helps if the wait returns spuriously or because the value changed. In this race, the wait blocks because the value is still expected, and no one will wake Thread A.

This is a well-known limitation of futex-style wait/notify when the notify does not change the value. The C++20 atomic_wait / atomic_notify design assumes notify is paired with a store that changes the value, but the proposal does not require this.

Files: include/wg14_atomic_waits/detail/impl/atomic_wait_linux.c.ipp:43-44, include/wg14_atomic_waits/detail/impl/atomic_wait_macos.c.ipp:64-66, include/wg14_atomic_waits/detail/impl/atomic_wait_common.ipp.ipp:494-499

4. Missing Tests

4.1 No tests for hash table path

The existing tests use atomic_int, which is 4 bytes. On Linux, macOS, FreeBSD, and Windows, 4-byte operations bypass the hash table. No tests exercise 1-byte, 2-byte, or 8-byte (Linux) operations that fall back to the hash table.

Impact: The critical bugs in sections 2.1, 2.2, and 3.1 are not caught by CI.

4.2 No tests for notify-without-store race

The tests always store a new value before calling notify. They do not test the scenario where notify is called without changing the value, which triggers the lost-wake race in section 3.2.

4.3 No tests for multiple waiters on the same object

atomic_notify_test.c creates multiple waiters but uses atomic_notify_all to wake them all. It does not test atomic_notify_one with multiple waiters, or the case where max_threads_to_wake limits the number of woken threads.

4.4 No tests for spurious wakes in hash table path

The tests do not inject spurious wakes into the hash table proxy to verify that the re-compare-and-re-park loop works correctly.

5. Other Issues

5.1 atomic_wait_expected_32 timeout conversion uses ceiling, but wait returns -1 on all errors

The timeout computation in atomic_wait_expected_32 correctly uses ceiling conversion (tv_nsec + 999999) / 1000000). However, because wait_on_address32 returns -1 for all errors (section 2.4), the caller cannot distinguish a real timeout from a spurious wake or other kernel error. The outer loop re-checks the deadline before each park, which mitigates this somewhat, but the return value semantics are still wrong.

Files: include/wg14_atomic_waits/detail/impl/atomic_wait_windows.c.ipp:52-53, include/wg14_atomic_waits/detail/impl/atomic_wait_macos.c.ipp:58

5.2 atomic_wait_common.ipp.ipp default proxy uses uint_least32_t but atomic prefix is missing

The default proxy definition:

#define WG14_ATOMIC_WAITS_HASH_TABLE_ITEM_PROXY_TYPE                           \
  WG14_ATOMIC_WAITS_ATOMIC_PREFIX atomic_uint_least32_t

The ATOMIC_PREFIX is applied, which is correct for C (atomic_uint_least32_t) but in C++ it becomes std::atomic_uint_least32_t. This is correct because the header includes <atomic> in C++ mode.

5.3 hash_table_remove_item doesn't reset proxy->atomic before free

Not a bug per se (the proxy is freed), but if the removal is ever skipped due to a non-zero use_count, the proxy remains with proxy->atomic = 1.

5.4 hash_table_grow doesn't reset old bucket keys to NULL after rehashing

In hash_table_grow, after rehashing, the old buckets are freed with free(table->buckets). The new buckets are freshly calloc'd, so they are zeroed. This is correct.

5.5 atomic_notify_32 fast-path for max_threads_to_wake == 0 returns 0 without CAS

The plan explicitly allows this:

"Use a fast-path check before acquiring the hash table lock: if max_threads_to_wake == 0, skip the hash table lookup and wake_waiters call entirely and return 0 immediately (no CAS, no notify, no side effects)."

Files: include/wg14_atomic_waits/detail/impl/atomic_wait_common.ipp.ipp:674-677

5.6 atomic_wait_generic sets order = success unconditionally after wait returns

If the wait returns because of an error (e.g., -EAGAIN from the proxy), order is still set to success. This means the subsequent load uses success ordering even though no notify occurred. The proposal requires failure ordering for accesses before any notify.

Files: include/wg14_atomic_waits/detail/impl/atomic_wait_common.ipp.ipp:419

5.7 atomic_wait_expected_32 return value on timeout is 0, but the proposal says "zero if no suspension occurred, or if duration timeout occurs"

This is correctly implemented: the function returns 0 on timeout and *expected is updated to the timed-out value.

Files: include/wg14_atomic_waits/detail/impl/atomic_wait_common.ipp.ipp:631-636

6. Summary of Critical Issues

The most severe issue is the hash table proxy state machine (section 2.1 / race 3.1). The proxy is a single-slot binary semaphore that is set to "notified" (1) by the notifier but never reset to "waiting" (0) after consumption. This makes the hash table path fundamentally broken for any object that receives more than one notify, or for any waiter that arrives after a notify has already been issued for that object.

Deepseek v4 Flash (0731) — Implementation Review

Date: 2026-08-01 Scope: Reviewed the wg14_atomic_waits implementation against docs/proposal.md (the authoritative spec) and plans/wg14_atomic_waits-reference-library.md. Method: Static inspection of every include/ backend, src/ wrapper, the public header, config.h, the tests, the build files, and CI. No debugger was run and no test program was executed; all conclusions come from reading the code.

1. Executive summary

The implementation is structurally faithful to the proposal (all six public APIs exist, width dispatch for 1/2/4/8-byte atomics, native wait/notify backends with a hash-table fallback), but it contains one primary correctness race that produces lost wakes (Section 2), a broken pthreads backend (Section 3), a timeout smuggled in as an error on Linux (Section 4), a Windows lost-wake for multi-thread wakeups (Section 5), plus a number of smaller deviations from the plan/proposal.

The single most important finding: the per-object "proxy" used by the hash-table path sets a 0/1 notification flag that is never cleared while the wait-queue node is alive. This is the lost-wake/livelock race the task asked to identify.

2. PRIMARY RACE: the proxy notification flag is never reset → lost wakes / busy-spin

Location

include/wg14_atomic_waits/detail/impl/atomic_wait_common.ipp.ipp:

• ..._WAIT macro (lines 37–38):

  wait_on_address32(&(x)->atomic, 0, (timeout))   /* wait while proxy->atomic == 0 */
  

• ..._WAKE macro (lines 39–45):

  atomic_store_explicit(&(x)->atomic, 1, release),  /* mono-directional: 0 -> 1 only */
  wake_by_address32(&(x)->atomic, max_threads_to_wake)
  

• atomic_wait_generic() (lines 327–435) — the shared "park by proxy" loop.
• atomic_notify_generic() (lines 437–456) — the shared "set flag + wake" path.

The bug

A waiter parks by calling WAIT(item,...) which blocks while item->atomic == 0 (i.e. FUTEX_WAIT(&item->atomic, 0)). A notifier sets item->atomic = 1 and wakes. There is no code anywhere that ever writes item->atomic back to 0 while the wait-queue node is alive. The only place it is reset is at node creation inside hash_table_find_or_create() (lines 220–222), which happens only when a brand-new proxy_waiter_t is allocated. A node is freed only when use_count drops to zero.

Consequence — two interleaving outcomes

(a) Re-park after a wake never sleeps (livelock). Once any notify has fired on a node, item->atomic is stuck at 1 for as long as the node lives. Any waiter that is woken and must re-park — the proposal explicitly requires re-park on spurious wake, and the code implements it as the top of the loop — calls FUTEX_WAIT(&item->atomic, 0) while the value is already 1. The kernel compares 1 != 0 and returns EAGAIN immediately. Every subsequent iteration of the loop returns immediately, so the waiter never sleeps again; it degenerates into a tight 100%-CPU busy-spin for the whole remaining lifetime of that node.

(b) Wakeups are lost because there is no sleeping thread. Because (a) means waiters stop sleeping, a later genuine producer store + notify_* sets the flag (already 1) and issues FUTEX_WAKE, which has nothing asleep to wake. The notify is therefore effectively lost for the purpose of the sleep/wake contract; correctness then depends entirely on the busy-spin poll observing the value change, which is not the semantics the proposal defines and not what a correct reference implementation should do.

Why the analogous futex idiom would be safe but this one is not

The correct pattern guards the "am I allowed to sleep" decision on the same state that the notifier flips, and the notifier re-arms the state before waiting:

• waiter: s = counter.load(); if (value == expected) futex_wait(&counter, s);
• notifier: counter++; futex_wake(...) Here counter is a strictly increasing generation so the waiter can always detect a change that happened between its load and its sleep. The implementation instead uses a single 0/1 flag that is never re-armed, so the invariant "atomic == 0 ⇔ a notify is pending/expected" is destroyed after the first notify.

Which configurations suffer

This path is the fallback for every backend whenever the operand cannot be handled directly by the kernel primitive, i.e. exactly the cases the proposal/plan force through the hash table:

• Linux: 1-, 2- and 8-byte operands (HAVE_WAIT_ON_ADDRESS_* is 32-bit only).
• macOS / FreeBSD: sub-native widths (1/2-byte).
• pthreads backend: every operand (there is no kernel per-address waiter).

The 4-byte Linux/macOS/FreeBSD/Windows fast paths and the atomic_wait_expected native-width path bypass the proxy and are not affected by this flag, but the 8-byte-on-Linux case — a perfectly legal and likely test target — is affected.

Recommended fix direction

Replace the 0/1 flag with a monotonically increasing sequence number that the waiter reads before parking and passes as the futex compare value, and that the notifier increments before waking. Reset-on-rearm must happen on the waiter side before the sleep decision, under the same lock used to re-check the object value (or rely on the kernel re-check for the object value itself as the futex fast path already does).

3. pthreads backend is fundamentally broken (hangs / lost wake)

include/wg14_atomic_waits/detail/impl/atomic_wait_pthreads.c.ipp:

• ..._WAIT (lines 45–46) is pthread_cond_wait(&(x)->atomic, pthreads_mutex()).
• pthreads_mutex() (lines 61–73) returns a _Thread_local mutex, i.e. a different mutex object per thread.

Problems:

1. pthread_cond_wait requires the passed mutex to be held by the calling thread. In atomic_wait_generic the waiter has released the hash-table lock (line 387) and then enters pthread_cond_wait with a mutex that is never locked. This is undefined behavior; on glibc it typically fails immediately (EPERM) so the wait “succeeds” without ever blocking — again a busy-loop — and there is no guarantee the node is protected.
2. The broadcast hand-off is not protected by the mutex the waiter sleeps on. A notifier holds the global hash-table lock and calls pthread_cond_signal (via the ..._WAKE macro, lines 47–53). The classic lost wake occurs when the notifier signals between the waiter’s re-check (value still equal to expected, line 380) and its pthread_cond_wait: the signal is dropped and the waiter blocks forever. With a futex, the kernel’s value re-check/EAGAIN saves this; with pthread_cond_t there is no such guard and there is no predicate/flag protecting the check, so the wait is a genuine, permanent lost wake (a hang).
3. Even the INIT/DESTROY macros treat pthread_cond_t through the generic proxy_waiter_t.atomic slot, but the shared atomic_wait_generic still performs flag-style logic (setting use_count, etc.) that is meaningless for a condvar.

Because CI runs ALWAYS_USE_PTHREADS_BACKEND=ON on Linux and macOS, this path is exercised, but its crashes/hangs are exactly the class of lost-wake bug being reported.

4. atomic_wait_expected mis-reports a timeout as an error on Linux

include/wg14_atomic_waits/detail/impl/atomic_wait_common.ipp.ipp, atomic_wait_expected_32() (lines 595–666), plus the Linux wait_on_address32() (atomic_wait_linux.c.ipp lines 37–52).

• The Linux wait_on_address32 returns 0 on success/EAGAIN/EINTR and -1 on any other error, not -errno. A genuine time-out of FUTEX_WAIT therefore comes back as -1 (with errno == ETIMEDOUT).
• The caller’s error branch (lines 651–659):

  if(ret2 < 0)
  {
    if(duration != NULL && ret2 != ETIME && ret2 != ETIMEDOUT) { errno = -ret2; return -1; }
  }
  

  ret2 is -1, which is never equal to the positive ETIME/ETIMEDOUT constants, so the condition is always true for any ret2 < 0 when a duration was supplied. A clean time-out returns -1 (error), not 0 (time-out) as the proposal requires:

  Returns: … returns zero … or duration timeout occurs.

This is timing-dependent — if the pre-wait clock_gettime check (lines 631–637) happens to notice expiry first it returns 0 cleanly — so the 1 ns test in atomic_wait_expected_test.c is flaky, but the underlying error path is wrong.

5. Windows wake_by_address* only ever wakes a single thread → lost wake

include/wg14_atomic_waits/detail/impl/atomic_wait_windows.c.ipp, wake_by_address32/wake_by_address64 (lines 92–124):

if(WakeByAddressSingle((PVOID)(uintptr_t) object)) return (max_threads_to_wake == 1) ? 1 : 1;
return 0;

• The max_threads_to_wake parameter is ignored; both atomic_notify_all and atomic_notify(..., max_threads_to_wake=N>1, ...) call this and wake exactly one thread via WakeByAddressSingle. The correct routine for max != 1 is WakeByAddressAll. Every other waiting thread is left parked → lost wake.
• The (max_threads_to_wake == 1) ? 1 : 1 ternary is dead code (both branches are 1).

This makes the Windows backend incorrect for atomic_notify_all and for atomic_notify with max_threads_to_wake > 1, which the plan marks as mandatory behaviour.

6. macOS timeout conversion deviates from the plan

atomic_wait_macos.c.ipp, wait_on_address32/64 (lines 55–68, 77–90):

• The plan (Step 11) requires: *duration → nanoseconds, cap each ulock_wait call at UINT32_MAX (~4.29 s) and loop for longer durations.
• The implementation instead converts once to microseconds (tv_sec * 1000000U + tv_nsec / 1000U) and passes it in a single call with no cap and no loop. For any duration ≥ ~4295 s the microsecond value overflows uint32_t, and durations beyond ~4.29 s are not split across multiple calls, so the accumulated wait can be far shorter than *duration — violating the proposal’s “total accumulated time … shall be at least *duration”.
• The code also declares private extern __ulock_wait/__ulock_wake instead of including <bsd/sys/ulock.h> as the plan directs; functional risk if SDK/version behaviour differs.

7. FreeBSD 8-byte UMTX_OP_WAIT argument-order inconsistency

atomic_wait_freebsd.c.ipp:

• 4-byte: _umtx_op(object, UMTX_OP_WAIT_UINT, expected, (long)&umtx_time) (lines 57–58) — passes the expected value in the value slot.
• 8-byte: _umtx_op(object, UMTX_OP_WAIT, (long)&umtx_time, (long)expected) (lines 88–89) — passes the timeout pointer in the value slot and the expected value in the address slot, i.e. the two are swapped relative to the 4-byte call.

The two calls are internally inconsistent, so at least one passes the operands in the wrong order; UMTX_OP_WAIT (8-byte) is almost certainly wrong and will compare against garbage / misbehave.

8. Return-value deviations from the plan

• atomic_notify_32 (lines 668–691) returns 1 + ret on a successful CAS, where ret is the number actually woken (0 on the proxy path when no node exists, 0 on the futex path when nothing is parked). So CAS-success-with-no-waiters returns 1 (positive), whereas the plan explicitly states:

  atomic_notify … Returns 0 if the CAS fails or no waiters are parked.

  (The proposal’s “possibly one plus the number woken” makes 1 defensible, so this is a plan deviation, not a proposal violation — noted for completeness.)

• atomic_wait_generic sets ret = 1 (line 388) before the first actual park call. The proposal ties “positive” to “suspended at least once”; the intent-to-park flag is acceptable but slightly loose.

9. Width-dispatch macros silently no-op on unsupported widths

atomic_wait.h, _WG14_ATOMIC_WAITS_IMPL_atomic_wait* / _notify* (lines 107–201):

• Each macro is if (sizeof==1) … else if (==2) … else if (==4) … else if (==8) … with no final else (and do{}while(0)). A _Atomic type of any other width (e.g. 16 bytes, or a long double, or a 0-width type) compiles to a silent no-op rather than a compile-time error, which can mask misuse. The plan states widths 1/2/4/8 are supported; a diagnostic would be safer.

10. Header-only / ODR notes

• The atomic_wait_*_N / atomic_notify*_N definitions in the .ipp files are not themselves marked WG14_ATOMIC_WAITS_INLINE/static (only their prior declarations in atomic_wait.h carry WG14_ATOMIC_WAITS_EXTERN, which is inline only when WG14_ATOMIC_WAITS_ENABLE_HEADER_ONLY). In the multi-TU header-only build (header_only_test*.cpp) this relies on the C inline extern-inline model. It is workable on GCC/Clang but fragile; the plan required every backend helper to be explicitly static inline to guarantee ODR safety.
• hash_table() uses WG14_ATOMIC_WAITS_IGNORE_MULTIPLE_DEFINITIONS (weak), so the singleton table is fine across TUs; this part is OK.

11. Smaller issues

• errno not preserved on success/failure paths of atomic_wait_expected_32: it writes errno = ETIMEDOUT (line 636) and errno = -ret2 (line 656), clobbering the caller’s errno even though errno preservation was a stated plan requirement (Steps 8–12). The futex wait_on_address32 does restore errno, but the caller overwrites it afterwards.
• EINTR treated as success (return 0) in wait_on_address32 (Linux line 45) rather than retrying; a signal during a timed wait silently shortens it.
• atomic_notify_generic drops a notify when no node is registered (find_or_create with increment_use_count=false, lines 444–445). For the futex backends this is masked by the waiter’s under-lock re-check; for the pthreads backend it is not masked (see Section 3), which is where the drop becomes a hard lost wake.
• Hash-table growth/deletion re-probes correctly (cluster-shift deletion, lines 271– 280), but hash_func truncates a 64-bit pointer to 32-bit before mixing (lines 108–113); not a correctness bug, just avoidable collisions.

12. Structure deviations from the plan (non-bug)

• The plan (Step 7) specified that atomic_wait/atomic_wait_explicit be thin wrappers delegating to atomic_wait_expected, and notify_one/notify_all delegate to a wake_waiters helper. The implementation instead uses per-width atomic_wait_1/2/4/8 + direct kernel calls for the native widths, with the hash-table path only for the fallback. This is a reasonable and arguably more correct choice given that atomic_wait_expected is native-width-only, but it is a deviation from the plan’s described structure.
• benchmark_atomic_wait_test.c does not actually run a benchmark (returns 0 after declaring an unused consumer), so CI excludes it correctly but it validates nothing.

13. Conclusion

The implementation gets the public API shape, the width dispatch, and the native-width futex path right, but the hash-table/fallback proxy logic has a one-way notification flag that is never re-armed, which is the race responsible for lost wakes: waiters that re-park on a node that has already been notified stop sleeping and busy-spin, and subsequent notifications have no sleeping thread to wake. On the pthreads backend the same logic is additionally broken by a thread-local mutex and a missing predicate under pthread_cond_wait, producing hard lost-wake hangs. These are compounded by a Linux timeout→error bug, a Windows single-wake bug for notify_all/max>1, a macOS timeout conversion that violates the minimum-duration guarantee, and a likely FreeBSD 8-byte argument swap. All were identified by code inspection only.