@classytic/primitives 0.28.0 → 0.30.0

package/CHANGELOG.md

@@ -1,1119 +1,1217 @@
-# Changelog
-
-Format based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
-adhering to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
-
-## 0.28.0
-
-### Added: a `pass` monetization kind (paid once, grants a term, never renews)
-
-The union could say "charged once" (`one_time`, a single price, no span) and
-"recurring" (`subscription`, priced spans that renew). It could not say
-"prepaid term": a course access pass, a gym or transit pass, a parking permit,
-a term licence. Hosts selling one reached for `subscription`, the only shape
-that carried priced spans, and then had to explain at every read that it does
-not actually renew. `toRevenueMonetizationType` would also have settled those
-sales as recurring income, forecasting revenue nothing was scheduled to bill.
-
-`PassMonetization { type: 'pass', passes: AccessPass[] }`, where an
-`AccessPass` is a key, a label, a price, an optional "was" price, and a
-`duration` + `durationUnit`. Shaped like `SubscriptionPlan` so a term means
-the same thing everywhere, kept separate because a plan carries `trialDays`
-and a renewal discount that a pass has none of.
-
-Also: `isPassMonetization`, `defaultPassOf` (the cheapest term, ties broken on
-the shorter SPAN so the headline is deterministic), and `approxTermDays` for
-ordering and labels only. `unitPriceOf` returns the cheapest term's price for
-a pass, `compareAtPriceOf` now returns the "was" of that SAME term so a card
-can never strike one term's price through beside another's, and
-`toRevenueMonetizationType('pass')` is `purchase`.
-
-A perpetual sale stays `one_time`: "forever" is the absence of a term, not a
-very long one.
-
-### Added — `CashDestination` on `./money`
-
-`CASH_DESTINATIONS` (`'cash' | 'petty_cash'`), `isCashDestination` and
-`cashDestinationLabel`: WHERE received cash landed, as an intent — never a
-chart-of-accounts code. One vocabulary for a request body, a domain event, a
-client SDK and a settlement dialog, so the accounting tier alone maps it to the
-active chart and a renumber cannot reach any of them. Added because a client
-carried a literal bank code the chart had already renumbered, and every
-settlement it sent was refused.
-
-## 0.27.0
-
-### Added — `./csv` subpath: one CSV cell escaper for the whole fleet
-
-`escapeCsvCell(value)` and `neutralizeCsvFormula(value)`. Six CSV producers
-across ledger, the spine and the hosts each hand-rolled RFC 4180 quoting, and
-not one neutralized a leading `= + - @ \t \r` — so a customer-supplied payment
-reference or party name reached the accountant's spreadsheet as a live formula.
-Quoting does not help: a spreadsheet strips the quotes and evaluates what is
-left, which is why the two defences ship as one module.
-
-The two functions are separable on purpose. A consumer holding a real CSV
-writer (`csv-stringify`) takes `neutralizeCsvFormula` only, so its quoting is
-not done twice; a hand-rolled `rows.map(...).join(',')` takes `escapeCsvCell`.
-
-**Well-formed numbers are exempt.** Prefixing `-1500` would make every negative
-amount text and break `SUM` in every finance export — a fix that quietly
-corrupts the product is worse than the hole. The exemption requires the WHOLE
-cell to be a number, so `-2+3+cmd|'/c calc'!A0` is still neutralized.
-
-Not for a machine-to-machine upload contract with a fixed column spec (a
-courier's bulk-import CSV): there the receiver is a parser, not a spreadsheet,
-and an added `'` corrupts the payload.
-
-## 0.26.4
-
-### Added — `./price-basis` subpath: measured pricing across physical dimensions
-
-New `PriceBasis` type (`FixedPriceBasis | MeasuredPriceBasis`) models how a quoted price spans a sellable quantity — flat fixed or per-unit-of-measure (mass, duration, volume, length, area). Factory functions `fixedPriceBasis()` and `measuredPriceBasis(dimension, qtyUnit, priceUnit, qtyPerPriceUnit)` construct the two shapes; `isPriceBasis()` is a structural guard for DB/wire boundaries. `resolvePriceBasisTotal(unitPriceMinor, quantity, basis?)` converts a minor-unit price × integer quantity to a final minor-unit total using half-even rounding (`roundedDiv` internally) and throws `PriceBasisError` on unsafe inputs or overflow.
-
-### Added — `./unit-cost-rate` re-exports `RoundingMode`
-
-`RoundingMode` (`'half-even' | 'half-up'`) is now defined in the new internal `rounded-div.ts` and re-exported from `unit-cost-rate`. No breaking change — existing consumers who import `RoundingMode` from `@classytic/primitives/unit-cost-rate` continue to work.
-
-### Changed — `validateAddress` / `isAddressValid` accept `Address` directly
-
-Both functions now accept `Address | PartialAddress | null | undefined`, removing the need for a cast when validating a fully-typed value object.
-
-## 0.26.3
-
-### Added — `PaymentProviderPort.wantsRawBody`
-
-Optional `readonly wantsRawBody?: boolean` on the payment gateway port.
-Providers that verify an HMAC signature (Stripe signs the bytes it transmitted)
-must declare `true` — the transport then hands the raw body rather than the
-parsed object. Absent / `false` preserves the historical behaviour (parsed
-body), which is correct for manual, cash and terminal providers that sign
-nothing.
-
-## 0.26.2
-
-### Added — `stockKeyFor` (`/item-identity`): one stock-key precedence for every package
-
-The product-level stock key, with a single precedence:
-
-1. the variant's stamped `skuRef` — assigned once at birth, immutable
-2. a **sole** variant with no stamp → the product's `_id`, never the label
-   (legacy writers derived `variantSku || productId`, and for one variant that
-   key IS the `_id`)
-3. a **sibling** with no stamp → its label, which genuinely is its legacy key
-4. nothing addressable → the product's `_id`
-
-The caller declares only what it HAS — `{ kind: 'sole' | 'label' | 'variant' }`
-— so the sole-vs-sibling decision is internal and a caller cannot make it wrong.
-That choice previously lived at every call site across three packages, in two
-functions with different fallbacks; they disagreed on a sole, unstamped,
-labelled variant, and the read then looked up a key no quant is stored under.
-
-It lives here, below every domain, because `@spinekit/order` and
-`@spinekit/catalog` must not import a peer domain — the rule sitting one layer
-too high is exactly why each package re-derived it.
-
-## 0.26.1
-
-- `./testing/stock-key-seam` — new subpath; `scanStockKeySeam`, `assertStockKeySeam`, `takesStockKeyFromLabel`, `STOCK_KEY_SEAM_FIXTURES` for conformance-scanning packages that resolve stock keys
-
-## [0.26.0] — 2026-08-27
-
-### Added
-
-- **`./deadline` subpath** — `Deadline` value type + helpers for date-bounded
-  obligations (reservations, SLA windows, contract terms).
-- **`./stock-level` subpath** — canonical stock-level representation and helpers.
-- **`./item-facts` subpath** — cross-kernel item fact contract (weight, dimensions,
-  hazmat flags) that catalog declares and logistics/inventory reads.
-- **`StockKey` branded type** (`./item-identity`) — prevents a bare merchandising
-  label or ObjectId being passed where a quant key is expected; the only source is
-  `stockKeyOf()` / `requireStockKey()`. Return types of both functions updated.
-- **Upfront collection policy** (`./monetization`) — `UpfrontCollectionPolicy`
-  (`'full_payment' | 'deposit' | 'deferred'`) + `UPFRONT_COLLECTION_POLICIES`
-  const. Cross-kernel: catalog declares the policy, order turns it into tenders,
-  revenue settles them.
-
-## [0.25.0] - 2026-08-24
-
-### Added
-
-- **`@classytic/primitives/item-identity` subpath** — `SkuFields`, `stockKeyOf`, `skuDisplay`. `sku` is a label; `skuRef` is the stock identity — the two fall back in opposite directions and getting one backwards type-checks while returning a plausible wrong string. `stockKeyOf` resolves `skuRef ?? sku ?? fallback` (the key to reserve, decrement and post against); `skuDisplay` resolves `skuLabel ?? sku ?? skuRef` (what a human reads). Canonicalises the per-package `variantSku || String(productId)` derivations into one place.
-
-## [0.24.0] - 2026-08-24
-
-### Added — `@classytic/primitives/address`: capture-format system
-
-`AddressFormat`, `AddressFieldDescriptor`, `AddressFieldKey`, `PartialAddress`, `AddressFieldError` — a description of HOW a country collects a `ContactAddress`: field labels, order, requiredness, subdivision taxonomy hooks, and postal code pattern. Country packs (e.g. `@classytic/bd-areas/address-format`) live beside their taxonomies; this subpath owns only the generic shape.
-
-`validateAddress(address, format)` — validates a partial address against a format; returns `AddressFieldError[]`. Enforces `required` on both canonical and extension fields identically, and checks `postalCodePattern` only when a value is given and a pattern is declared. `isAddressValid` and `describeAddressError` are convenience wrappers.
-
-Every `key` in a descriptor is a field of `ContactAddress`; co-location with the value object is what makes a mismatch obvious instead of invisible.
-
-## [0.23.0] - 2026-08-15
-
-### Added — `./approval`: `notRequiredChain(reason)` + `isNotRequired(chain)`
-
-The chain for a subject that needs **no** approval — zero steps, already resolved. The
-explicit, greppable answer to "the policy engine ran and decided this document does not
-require approval", which callers previously had no way to express.
-
-This states a conclusion the module already reached rather than inventing one:
-`computeChainStatus` folds an empty step list to `'approved'` — "nothing outstanding" is
-what approved MEANS here, and that fold predates this function. `createChain` still throws
-`EMPTY_STEPS`: building a ROUTING chain that routes nowhere is a different mistake and
-stays refused. The two are deliberately not interchangeable, so reaching "approved with no
-approvers" requires saying so by name.
-
-**Why a chain rather than leaving the field unset.** Absence is the one input that occurs
-by accident — a schema that never declared `approvals` strips it on write — and
-`assertApproved` must keep treating that as `CHAIN_MISSING`. A present zero-step chain is
-distinguishable from a lost chain (absent) *and* from a granted one (has steps and
-decisions), so existing gates keep working through `isApproved` alone. No consumer needs a
-second predicate, which is what would drift.
-
-`isNotRequired` tests the STRUCTURE (`steps.length === 0`), not the new
-`notRequiredReason` field: a subject schema that never declared that field drops it on
-write, so a predicate reading it would answer differently before and after the save. The
-reason is provenance for the audit trail only — pinned by a test that strips it and asserts
-the predicate is unmoved.
-
-**The hazard, stated in the docblock:** this makes `isApproved` true with no human
-decision. It is correct only where a policy engine ASKED and got "no policy applies" —
-never as a fallback for a lookup that failed or was never configured. First consumer:
-`@classytic/arc-approval` 0.4.0 `whenNoPolicy: 'not-required'`.
-
-## [0.22.0] - 2026-08-12
-
-### Added — outbox fencing tokens (`claimPendingFenced`)
-
-- **`OutboxStore.claimPendingFenced?`** (additive, feature-detected — mirrors repo-core's `LockAdapter.tryAcquireFenced`): same atomic semantics as `claimPending`, returning `OutboxClaimedEvent { event, fencingToken }`. The token is minted by the STORE per claim epoch — a takeover after lease expiry mints STRICTLY GREATER, so a stale ex-holder's `acknowledge`/`fail` (now accepting `fencingToken?`) throws `OutboxOwnershipError`. This closes the overlap that consumer-id ownership checks narrow but cannot close: holder ids recur; epochs do not.
-- `MemoryOutboxStore` implements it; `claimPending` delegates to the fenced path (one claim implementation). Stale-ack rejection is mutation-proven.
-
-### Added — `./period`: `shiftPeriod(period, n)` — calendar arithmetic without instants
-
-Shift a MONTH or QUARTER period by `n` of its own units (negative = back); YEAR shifts by years; a `custom` (ad-hoc start/end) period has no unit to shift and throws rather than guessing one. No `Date` is constructed and no zone is consulted — shifting a month is pure `(year, month)` arithmetic, and the `timezone` rides along unchanged so the shifted period still resolves against the calendar it was defined in.
-
-Exists because both inline shapes callers reach for are wrong the same way: `new Date(Date.UTC(y, m - 1 - n, 1))` answers a business-calendar question with a UTC instant (six hours off at both ends in Asia/Dhaka — the bug class this module's header documents five instances of), and hand-rolled `y * 12 + (m - 1) - n` integer math is a third copy of a rule `Period` already owns, with the year rollover exactly where copies go wrong. First consumer: be-prod `business-date.ts` ("two filing months ago").
-
-**Versioning note:** this export briefly existed in a working copy labelled 0.21.0 after 0.21.0 had already been published without it — same version, different surface. 0.22.0 restores the invariant that a version number names exactly one contract; `>=0.22.0` is the correct floor for any consumer importing `shiftPeriod`.
-
-### Docs — `./environment` docblock refinement (no API change)
-
-The 0.21.0 entry below originally listed `isProduction` / `isDevelopment` / `isTest`; the published surface is `classifyEnv`, `EnvClass`, `isProductionEnv` only — `isProductionEnv` is deliberately the single predicate (consumers needing all three buckets build their own triple from `classifyEnv`). The entry below is corrected in place; runtime and declarations are byte-identical to published 0.21.0 apart from comments.
-
-## [0.21.0] - 2026-08-10
-
-### Added — `./environment`: canonical `NODE_ENV` classification
-
-`classifyEnv(raw)` — total function mapping a raw `NODE_ENV` string to `EnvClass = 'development' | 'test' | 'production'`. Every input yields exactly one bucket; no fourth state or `undefined`. Handles both spellings per bucket (`production`/`prod`, `test`/`qa`), trims whitespace (`.env` files and CI variables routinely carry trailing newlines), and defaults unknown values to `development` — the safe-fail direction that keeps local tooling functional.
-
-Collapses four copies across the codebase that each missed at least one spelling: `be-prod classifyEnv`, `gym businessMode()`, arc's CLI config template (`z.preprocess`), and arc's four raw comparison sites. The missed spellings were all in the dangerous direction — a `NODE_ENV=prod` instance that matched nothing fell through to `development`, opening security gates (`Secure` cookie flag, arc's dev preset, non-transactional money writes, raw 500 messages to the client) in production.
-
-Exports: `classifyEnv`, `EnvClass`, `isProductionEnv`.
-
-### Added — `publint` gate on build
-
-`publint: true` in `tsdown.config.ts` — catches export-map entries pointing at non-existent dist files at build time rather than at consumer install time.
-
-## [0.20.0] - 2026-08-05
-
-### Added — `./period` OWNS timezone-aware period resolution (was two copies above the kernels)
-
-`Period` was `{ year, month?, quarter? }` with no timezone and no resolver, so it
-could not express "the business month of August in Asia/Dhaka". The resolver
-existed TWICE, both ABOVE the kernels — `@spinekit/kit/period`
-(`createBusinessCalendar`) and be-prod's `#lib/utils/business-date` — which meant
-a kernel could reach neither and re-derived the boundary inline with `Date.UTC`.
-Five silent UTC-month defects came out of that (`aggregateMonthlyVat`, the
-sales-fact reconciler, and three in bd-tax Mushak).
-
-This is a MOVE of the proven `createBusinessCalendar` logic, not a re-derivation:
-same civil-date round-trip through `/timezone`, same half-open ranges, same
-DST-exactness.
-
-- **`Period.timezone?`** — the IANA zone a period is DEFINED in, so a persisted or
-  transported period keeps its own calendar.
-- **`resolvePeriod(period, timezone?)`** → half-open `DateRange`. Throws when no
-  zone is available anywhere (never silently answers in UTC) and throws when the
-  period's zone and the caller's zone DISAGREE.
-- **`resolveDay` / `resolveMonth` / `resolveQuarter` / `resolveYear` /
-  `resolveDateSpan` / `dayStart`** — the granularity-specific forms. All return
-  HALF-OPEN `[start, end)`, which tiles exactly.
-- **`periodOf(instant, granularity, timezone)`** — the inverse: which filing
-  period an instant belongs to on the LOCAL calendar.
-- **`parsePeriod` / `formatPeriod` / `granularityOf` / `periodTimeZone`** —
-  `'YYYY'` · `'YYYY-MM'` · `'YYYY-Qn'`. `parsePeriod` REFUSES a `'YYYY-MM-DD'`
-  rather than reading it as its month (a ~30× widening).
-- **`inclusiveEnd(range)`** — `end - 1ms` for legacy `$lte` call sites, derived
-  from the resolved end so it stays DST-exact.
-- **`PeriodError`** + **`PeriodErrorCode`** (`AMBIGUOUS_PERIOD` ·
-  `INCOMPLETE_RANGE` · `INVALID_FIELD` · `MISSING_TIMEZONE` · `TIMEZONE_CONFLICT`
-  · `INVALID_LABEL`).
-
-A `Period` carrying two resolution shapes (`month` + `quarter`, or either with
-`start`/`end`) now throws instead of preferring one.
-
-### Added — `./idempotency`: the claim + lease CONTRACT
-
-Three kernels hand-rolled incompatible versions of one state machine — `cart`
-(`in_flight`/`succeeded`/`failed`, the only one with crash-window leasing),
-`catalog` offers (`pending`/`completed`, no lease), `contract` amendments
-(implicit). Types + PURE helpers only; no persistence, no collection, no driver.
-
-- **`ClaimOutcome = 'claimed' | 'replayed' | 'in_flight'`** — three-valued, and
-  `in_flight` is NOT a failure. Mapping a live concurrent lease onto an error
-  licenses a retry, which is a double-apply.
-- **`IdempotencyClaim`** (`identity` · `requestFingerprint` · `state` ·
-  `leaseToken` · `leaseExpiresAt` · `attempts` · `result`), **`ClaimResult`**,
-  **`ClaimState`**, **`IdempotencyIdentity`**.
-- **`decideClaim(existing, request)`** — the whole state machine as one pure
-  function. Throws `FINGERPRINT_MISMATCH` (same key, different body) and
-  `MISSING_REPLAY` (terminal record with no stored result) rather than answering.
-- **`completeClaim` / `renewLease`** — both refuse when the caller lost its lease,
-  so a superseded attempt cannot overwrite the winner's result.
-- **`identityKey`** (canonical-JSON composite, so a delimiter inside a scope value
-  cannot collide two identities), **`fingerprintRequest`**, **`newLeaseToken`**,
-  **`isLeaseExpired`**, **`holdsLease`**, **`isInFlight`**, **`DEFAULT_LEASE_MS`**.
-
-### Changed — BREAKING: five API decisions taken before spine wiring starts
-
-- **`./cadence` — `Cadence.timezone` is now AUTHORITATIVE, not display-only.**
-  It was documented as decorative while `stepMonthly`/`stepYearly` used
-  `getUTCMonth()`/`getUTCDate()`, so `timezone: 'Asia/Dhaka'` + `dayOfMonth: 1`
-  with a midnight anchor placed every occurrence on the local **2nd**. All four
-  kinds now run on the zone's wall clock when one is set (daily advances LOCAL
-  days — 23h/25h across a DST transition), and stay UTC when it is absent.
-  `validateCadence` rejects an unresolvable zone with the new
-  `INVALID_TIMEZONE`. **Behavioural change for any persisted cadence that
-  already carries a `timezone`** — the occurrences move to where the field
-  always claimed they were.
-- **`./approval` — `assertApproved` no longer passes on a `null` chain.** It now
-  throws the new `CHAIN_MISSING`. Absence is not approval, and a mongoose schema
-  that never declared `approvalChain` strips it on write, so the gate could not
-  tell "no approval configured" from "approval silently lost". The explicit
-  opt-out is the new **`assertApprovedIfPresent`**. The second parameter is now
-  an options object (`{ message }`), not a bare string.
-- **`./payment-events` — every payload carries an `eventType` discriminant, and
-  there is a `payment.unknown` event.** `PaymentInitiatedPayload` and
-  `PaymentSucceededPayload` were structurally IDENTICAL with no discriminant on
-  the union, so a handler typed for "funds received" accepted "money in flight".
-  Adds **`PaymentUnknownPayload`** (the port contract's third outcome had no
-  event, so an unobserved result had to be forced into SUCCEEDED or FAILED),
-  **`PaymentEventPayloadMap`**, **`isPaymentEvent`**, **`isFundsReceived`**, and
-  three compile-time drift guards over the map.
-- **`./currency` — `convertWithSnapshot` / `reverseWithSnapshot` take and return
-  `Money`.** A bare `number` meant nothing checked the amount was denominated in
-  `fx.sourceCurrency`; a USD→BDT snapshot applied to a EUR amount type-checked
-  and returned a plausible wrong number. Both now throw `CurrencyMismatchError`
-  on a mismatch, handle differing minor-unit exponents (USD 2 → JPY 0), and take
-  an explicit **`FxRounding`** (`'half-away-from-zero'` default · `'floor'` ·
-  `'ceil'`) instead of returning an unrounded float for each caller to round
-  differently.
-- **`./money` — `Money.currency` is `CurrencyCode`, not `CurrencyCode | string`.**
-  A union with `string` collapses to `string`, so the brand was decoration on
-  the one type where it matters and `money()` validated nothing. `money()`,
-  `fromMajor()` and `sumMoney()` now validate and brand; `isMoney` format-checks
-  the code instead of `typeof === 'string'`. New **`currencyCode(value)`**
-  (throwing constructor) and **`InvalidCurrencyCodeError`**.
-  **`CurrencyMismatchError` is now DECLARED in `./currency`** (so the FX seam can
-  throw it without an ESM cycle) and re-exported from `./money`, so existing
-  imports are unaffected.
-
-Migration: a `Money` object literal now needs a branded code —
-`money(1999, 'USD')`, or `{ amount: 1999, currency: currencyCode('USD') }`.
-
-### Changed — six silent-permissiveness defects now FAIL instead of answering
-
-Every item below previously returned a plausible-looking value where it should
-have errored. All are behaviour changes for malformed input only; correct input
-is unaffected. Falsifying tests live in `tests/unit/fail-loud-regressions.test.ts`.
-
-- **`./period` — an Invalid Date no longer reports "no overlap".** `rangesOverlap`,
-  `isWithin` and `rangeDurationMs` throw the new `DateRangeError` on a non-finite
-  endpoint; `isDateRange` returns `false` for one. Every comparison against `NaN`
-  is `false`, so a corrupt range used to answer "no conflict" — the PERMISSIVE
-  answer for the three documented consumers (order booking collision, flow
-  reservation-window collision, promo campaign overlap), i.e. a silent double-book.
-- **`./currency` — the ISO 4217 exponent table is now exhaustive.** All 0-decimal
-  (`BIF CLP DJF GNF ISK KMF KRW PYG RWF UGX UYI VND VUV XAF XOF XPF`), 3-decimal
-  (`BHD IQD JOD KWD LYD OMR TND`) and 4-decimal (`CLF UYW`) currencies are listed,
-  so `minorUnitFactor`'s `?? 100` fallback is reachable only by a code that is not
-  a real ISO currency. Previously `fromMajor(1500, 'GNF')` stored 150 000 minor
-  units — a 100× error that looks like a number. New `isKnownCurrency(code)`
-  distinguishes "listed" from "assumed" for boot-time config validation.
-- **`./proration` — a missing allocated part throws instead of becoming `0`.** The
-  `?? 0` fallbacks in `splitByPeriodFraction` / `allocateMoneyByFraction` are now
-  `ProrationError('ALLOCATION_INVARIANT')`. A zero credit and a genuinely
-  fully-consumed period were indistinguishable downstream.
-- **`./state-machine` — `assertAndClaim` rejects an EMPTY `from` list.**
-  `validSources(to)` returns `[]` for an unreachable status, and the documented
-  `from: machine.validSources(x)` idiom then ran the assert loop zero times and
-  handed the CAS an empty allow-list. An empty allow-list must never mean "any".
-- **`./timezone` — the offset regex is anchored.** An offset label the runtime
-  emits in an unexpected shape (e.g. an unpadded `GMT+5:30`) matched the optional
-  bare-`GMT` branch and resolved to **UTC**, silently shifting every business-day
-  boundary derived from it. Unrecognised labels now throw; `\d{1,2}` accepts the
-  unpadded variant rather than mis-reading it.
-- **`./canonical` — class instances are rejected, not collapsed.** A
-  `Types.ObjectId`, `Buffer` or hydrated Mongoose subdocument has no own
-  enumerable keys, so it digested as `{}` and two manifests referencing DIFFERENT
-  documents hashed identically — the integrity check reported "unchanged".
-  `canonicalJson` now throws `CanonicalizeError` for any object whose prototype is
-  neither `Object.prototype` nor `null`. Plain objects, null-prototype objects,
-  arrays and `Date` are unaffected.
-
-### Changed — the publish smoke gate now sweeps the whole export map
-
-`tests/smoke/smoke.mjs` derives its subpath list from `package.json#exports`
-(44/44) instead of a hand-maintained list that covered 23, and asserts each entry
-declares both `types` and `default` and that both files exist. The gate runs in
-`prepublishOnly`; previously a subpath that lost its tsdown entry would have
-published an export map pointing at a missing file with the gate printing OK.
-
-## [0.19.0] - 2026-08-01
-
-### Added — `./monetization`: canonical pricing classification contract
-
-- **`@classytic/primitives/monetization`** — the single home for how a product
-  is sold and priced, so the classification survives end-to-end across catalog,
-  revenue, order, and entitlement kernels without each maintaining its own enum.
-- **`Monetization`** discriminated union — `FreeMonetization | OneTimeMonetization |
-  SubscriptionMonetization | BundleMonetization | UsageMonetization`, each
-  carrying its own nested pricing shape.
-- **`MONETIZATION_KINDS`** + **`MonetizationKind`** + **`isMonetizationKind`** —
-  exhaustive kind list with runtime guard.
-- Type-guard predicates: **`isFreeMonetization`**, **`isOneTimeMonetization`**,
-  **`isSubscriptionMonetization`**, **`isBundleMonetization`**,
-  **`isUsageMonetization`**.
-- Supporting types: `PriceTier`, `OneTimePricing`, `SubscriptionPlan`,
-  `UsageTier`, `UsageRating`, `DurationUnit`, `MoneyByCurrency`.
-
-### Added — `./proration`: pure proration arithmetic
-
-- **`@classytic/primitives/proration`** — timezone-agnostic, policy-free
-  proration math. Holds no commercial policy (no upgrade/downgrade semantics,
-  no discount or tax logic) — those live in the contract kernel.
-- **`periodProgress(input)`** — fraction of a billing period consumed given
-  start, end, and current instants as `Date`. Returns `PeriodFraction`.
-- **`splitByPeriodFraction(amount, fraction, granularity?)`** — split a `Money`
-  amount into consumed/remaining parts; `granularity: 'whole_day'` (default) or
-  `'exact'`.
-- **`allocateMoneyByFraction(amount, fraction)`** — proportional allocation
-  using the largest-remainder allocator so `consumed + remaining === amount`
-  exactly (no penny drift).
-- **`ProrationError`** + **`ProrationErrorCode`** (`INVALID_PERIOD` |
-  `INVALID_FRACTION`) — typed errors for out-of-range inputs.
-- Supporting types: `PeriodFraction`, `PeriodProgressInput`,
-  `ProrationGranularity`.
-
-## [0.18.0] - 2026-07-29
-
-### Added — `./subject`: `SubjectRef` polymorphic identity pointer
-
-- **`@classytic/primitives/subject`** — `SubjectRef` (`{ subjectModel, subjectRef }`) — WHO a record is about, polymorphically. The pairing spans every kernel that holds data about a person without owning them (attendance, access, loyalty, credentials). Deliberately kept distinct from `ExternalRef` (provenance — "came from") even though the shapes are identical: a grant carries both at once and they mean different things.
-- **`buildSubjectRef(model, id)`** — returns `SubjectRef | null`; `null` when `id` is absent/blank, because a malformed binding silently matches nothing (an entitlement nobody holds, a credential no revoke can find) rather than carrying `"undefined"` as a string.
-- **`SubjectRef`** type exported from `@classytic/primitives/subject`.
-
-### Added — `./state-diagram`: Mermaid diagram from a `StateMachine`
-
-- **`@classytic/primitives/state-diagram`** — `renderStateDiagram(machine, opts?)` — generates a `stateDiagram-v2` Mermaid string from the same `StateMachine` object the runtime CAS enforces. Removes the class of drift where a hand-drawn prose comment diverges from the actual transition table. Pure: string in, string out, no I/O.
-- **`StateDiagramOptions`** — `title?`, `highlight?` (mark specific states), `noteOn?` (per-state annotations).
-
-### Added — `./suspension`: time-bounded suspension/pause policy
-
-- **`@classytic/primitives/suspension`** — pure suspension policy computation (no persistence). Models a pause the subject is ENTITLED to take, bounded by an annual allowance, that ends by itself — distinct from `/hold` (an indefinite blocker needing manual resolution).
-- **`evaluateSuspension(policy, history, opts)`** — always returns `autoResumeAt` (the instant the pause MUST end) so a sweep can resume it whether or not the subject ever comes back. Covers: gym membership freeze, subscription pause, unpaid leave.
-- **`SuspensionPolicy`** — `maxDaysPerPeriod`, `minDurationDays`, `period` ('annual' | 'rolling').
-- **`SuspensionEntry`** — a historical pause record (`startedAt`, `endedAt?`, `reason?`).
-- **`SuspensionResult`** — `autoResumeAt`, `daysUsed`, `daysRemaining`, `status`.
-
-### Added — `OutboxStore.transactionalSave` (`/outbox`)
-
-- **`OutboxStore.transactionalSave?: boolean`** — declares whether `save` enlists `options.session` in the caller's transaction. Absent/`false` = "assume NOT atomic". Set `true` ONLY when `save` passes the session to the same transactional resource the caller writes (e.g. a mongoose model on the same connection). Callers that need atomicity (`@classytic/access`'s `atomically()`) check this flag and fail at boot rather than claiming a guarantee they cannot keep.
-- **`MemoryOutboxStore.transactionalSave = false`** — explicitly declared; an in-memory store cannot participate in a Mongo transaction, so its write survives a rollback.
-
-### Changed — `shelf-life.ts` deduplicates `addDays`
-
-- `addDays` is now imported from `./calendar` instead of being re-defined inline. No behavior change.
-
-## [0.16.0] - 2026-07-27
-
-### Added — `assertApproved` + `CHAIN_INCOMPLETE` (`/approval`)
-
-- **`ApprovalErrorCode` gains `'CHAIN_INCOMPLETE'`** — the single canonical
-  error code for "finalize/commit attempted before the chain reached `approved`".
-  Hosts route it through `rethrowApprovalError` to the `approval.chain_incomplete`
-  wire code so all commit/post/finalize boundaries raise the same structured error.
-- **`assertApproved(chain, message?)`** — canonical finalize-time gate: asserts
-  the chain has reached `approved`, else throws `ApprovalError('CHAIN_INCOMPLETE')`.
-  A `null`/absent chain passes (no chain required). Use at every journal-entry
-  post, PO approve, transfer dispatch, etc. to replace per-domain hand-rolled
-  divergent codes with one definition.
-
-Purely additive — no existing exports changed.
-
-## [0.15.0] - 2026-07-25
-
-### Added — `/retention`: `PurgeEvidence` value object
-
-- **New `@classytic/primitives/retention` subpath** exporting **`PurgeEvidence`**
-  — the persisted evidence record for a GDPR / data-retention purge operation
-  (proof of what was purged, over what scope, by whom, why, and whether
-  analytical measures were retained). Shape: `{ id, subject: { ref, model? },
-  scope, strategy: 'hard'|'soft'|'anonymize', measuresRetained, processed,
-  occurredAt, actor: { ref, kind: 'user'|'system'|'service' }, reason,
-  legalBasis?, status: 'completed'|'partial'|'failed', operationId?,
-  startedAt?, completedAt?, results?: PurgeResourceResult[],
-  verification?: PurgeVerificationSummary }`.
-- **`createPurgeEvidence(input)`** — pure builder (auto-fills `id` + `occurredAt`
-  when absent, defaults `status` to `'completed'`). **`isPurgeEvidence(value)`**
-  type guard + **`assertPurgeEvidence(value)`** throwing assertion.
-
-### Added — `/canonical`: strict deterministic JSON + integrity digest
-
-- **New `@classytic/primitives/canonical` subpath** — `canonicalJson(value)`
-  produces a deterministic string (keys sorted recursively, `Date` → `{"$date":"<iso>"}`,
-  strict rejection of `undefined`/`NaN`/`BigInt`/`Map`/`Set`/cycles with
-  `CanonicalizeError`). **`sha256Hex(input)`** and **`canonicalDigest(value)`**
-  (= `sha256Hex(canonicalJson(value))`) — integrity checksums, not tamper-proof
-  seals. Uses `node:crypto` directly (same precedent as `/otp`).
-
-### Added — `/event-infra`: `propagateHandlerErrors` option
-
-- **`InProcessEventBusOptions.propagateHandlerErrors?: boolean`** (default
-  `false` — existing error-isolating behavior unchanged). When `true`, `publish`
-  runs handlers sequentially and rethrows the first handler error to the
-  publisher — for single-consumer projection/relay lanes where a failed handler
-  must not be silently acked. `publishMany` maps each event's failure into its
-  `PublishManyResult` entry instead of rejecting the batch.
-
-## [0.13.0] - 2026-07-22
-
-### Added — `/unit-cost-rate`: exact per-unit monetary rates (bigint-safe)
-
-- **`@classytic/primitives/unit-cost-rate`** — fractional per-unit monetary
-  rates stored as scaled integers (`scaledAmount = minorPerUnit × RATE_SCALE`,
-  `RATE_SCALE = 1_000_000`). Solves the WAC / FIFO valuation problem where
-  dividing a total by a fractional quantity yields a sub-minor-unit rate
-  (100 paisa ÷ 3 units = 33.333… paisa/unit) that must not lose precision.
-  Exports:
-  - `unitCostRate(minorPerUnit, currency)` — direct rate constructor
-  - `unitCostRateFromTotal(totalMinor, quantity, currency)` — WAC/FIFO
-    constructor; bigint-safe, QTY_SCALE grid for fractional quantities
-  - `extendedAmount(rate, qty, mode?)` — the ONE rate→total rounding boundary;
-    bigint multiplication to keep large amounts exact; default `'half-even'`
-    (banker's rounding), `'half-up'` available for legacy parity
-  - `rateMinorPerUnit(rate)` — display helper (fractional number, NOT for math)
-  - `isUnitCostRate(value)` — structural type guard
-  - `UnitCostRate<C>` interface, `UnitCostRateError`, `UnitCostRateErrorCode`,
-    `RoundingMode`, `RATE_SCALE`
-  - For distributing a top-down charge across lines with no lost minor unit,
-    compose with `allocate` from `@classytic/primitives/split-allocation`.
-
-### Changed — `/outbox` is now the canonical contract owner
-
-- **`@classytic/primitives/outbox` owns the transactional-outbox contract.**
-  Previously this file was documented as a mirror of `@classytic/arc`'s copy
-  ("arc is source of truth, keep bit-identical") while arc's docs said the
-  opposite — a manual-sync drift risk. Settled: primitives owns pure
-  cross-package contracts (events + outbox); arc owns runtime behavior
-  (`EventOutbox` relay, stores, backoff, Fastify integration) and, from
-  arc 2.24+, imports + re-exports this contract instead of duplicating it.
-- **Cross-package `instanceof` now works.** Because arc re-exports these exact
-  `OutboxOwnershipError` / `InvalidOutboxEventError` classes, a store built
-  against primitives throws the same class identity arc's relay catches.
-  Under the old duplication, arc's `instanceof` checks silently missed
-  primitives-thrown ownership errors.
-- No export names or signatures changed — contract surface is identical.
-
-## [0.11.0] - 2026-07-08
-
-### Added — eKYC vocabulary + secure OTP primitive
-
-- **`@classytic/primitives/identity-verification`** — cross-package eKYC data shapes:
-  `IdentityDocumentType`, `VerificationCheck`, `VerificationLevel`, `SubjectKind`,
-  `IdentitySubject`, `DocumentReference`, `IdentityProviderCapabilities`,
-  `VerificationStatus`, `CheckResult`, `CreateVerificationParams`,
-  `VerificationSession`, `VerificationWebhookEvent`. Pure helpers: `isVerified`,
-  `checksSatisfy`, `deriveStatus`. The identity twin of `payment-gateway` — shapes
-  here, contracts in engines, provider implementations anywhere. Serves `@classytic/esign`
-  (IdentityBridge port), `@classytic/payee` (ComplianceBridge), and PSP merchant KYC.
-- **`@classytic/primitives/otp`** — secure OTP challenge lifecycle. Guarantees: code
-  never persisted (only HMAC-SHA256 digest stored), constant-time `timingSafeEqual`
-  comparison, rejection-free `randomInt` CSPRNG (no modulo bias), bounded attempt
-  counter, TTL expiry, single-use verified state. Exports: `createOtpChallenge`,
-  `verifyOtpChallenge`, `isOtpVerified`, `isOtpActive`. Delivery stays in
-  `@classytic/notifications` — this primitive owns only the verify half.
-
-## [0.9.1] - 2026-07-04
-
-### Added
-
-- **`CURRENCY_PATTERN` exported from `/currency`** — the ISO 4217 regex
-  (`/^[A-Z]{3}$/`) behind `isCurrencyCode` / `toCurrencyCode`. Consumers
-  embed the SAME pattern in JSON-Schema-representable validators (zod
-  `.regex`, Mongoose `match`, OpenAPI `pattern`) instead of hand-rolling
-  copies — `.refine(isCurrencyCode)` validates at runtime but is dropped
-  by `z.toJSONSchema`, silently losing the constraint from generated API
-  docs.
-
-## [0.9.0] - 2026-07-02
-
-### Added — `/timezone`: IANA resolution + civil dates (the door for `/calendar`'s escape hatch)
-
-`/calendar` owns wall-clock ARITHMETIC over a fixed `UtcOffsetMinutes` and
-tells hosts to "resolve the offset for the instant and pass it in" — but the
-stack had no shared resolver, so consumers hand-rolled `Intl.DateTimeFormat`
-(three divergent copies found: catalog ×2, be-prod POS). `/timezone` is that
-resolver — zero-dep (Intl IS the ICU tz database), Temporal-shaped names for
-a future mechanical migration:
-
-- **`zoneOffsetMinutes(instant, zone)`** — per-instant, DST-exact offset;
-  composes directly with `/calendar`: `startOfMonth(i, zoneOffsetMinutes(i, z))`.
-- **`localTimeParts(instant, zone)`** — `{ isoWeekday, hour, minute,
-  minutesOfDay }` in the zone; what pricing windows / slot grids /
-  working-hours checks evaluate against (canonicalizes catalog's resolvers).
-- **`CivilDate`** (branded `'YYYY-MM-DD'`) + `civilDate` / `isCivilDate` /
-  `civilDateOf(instant, zone)` / `civilDateToInstant(cd, zone, time?)`
-  (two-pass DST-safe) / `addCivilDays` / `civilDaysBetween` — the type for
-  hotel nights, POS business dates, VAT filing years. Lexicographic ==
-  chronological, so it's a range-queryable Mongo string key.
-- **`isValidTimeZone(zone)`** — boot-time host-config validation;
-  **`listTimeZones()`** + **`zoneOffsetLabel(instant, zone)`** — the data
-  source + label for Google-Calendar-style zone pickers, straight from ICU
-  (`Intl.supportedValuesOf`), no bundled list to go stale.
-- `TimeZoneError`; formatter instances cached per zone.
-
-No arithmetic here (that stays in `/calendar`) — resolution only.
-
-## [0.7.2] - 2026-05-30
-
-### Added — `BankTransaction` per-account routing + pending lifecycle fields
-
-Three additive optional fields on `BankTransaction` (`/bank-transaction`):
-
-- **`sourceAccountId?: string`** — the vendor-side account this row belongs to
-  (Plaid `account_id`, Xero `BankAccount.AccountID`). Sync feeds return
-  transactions spanning multiple accounts in one batch; this lets consumers
-  route each row to the right account per-transaction (multi-account sync)
-  instead of forcing the whole batch onto one account. Statement formats
-  (single-account files) leave it undefined — the account lives on
-  `BankStatement.account`.
-- **`pending?: boolean`** — the entry hasn't posted/cleared yet (real-time
-  feeds like Plaid `pending: true`; statement formats leave it undefined).
-  Consumers must treat pending rows as provisional — don't reconcile/post
-  them as final.
-- **`supersedesExternalId?: string`** — set on a POSTED row that replaces an
-  earlier PENDING one, carrying the pending row's `externalId` (Plaid
-  `pending_transaction_id`). Pending and posted rows have different
-  `externalId`s, so consumers MUST use this back-pointer to drop/replace the
-  superseded pending row on ingest — otherwise both persist as duplicates.
-
-Backward-compatible (both optional). Closes a fidelity gap where Plaid's
-pending state was dropped at the canonical boundary, so downstream consumers
-(`@classytic/fin-io` Plaid mapper, `@classytic/revenue` import) couldn't
-distinguish provisional rows or dedupe pending→posted.
-
-## [0.7.1] - 2026-05-26
-
-### Added — workflow + scheduling + composition primitives
-
-Four new flat subpaths (additive):
-
-- **`/status-history`** — append-only status transition log primitive.
-- **`/condition`** — declarative condition expressions for workflow gates.
-- **`/mixin`** — composition helpers for primitive object assembly.
-- **`/sla-policy`** — SLA policy definitions paired with the existing `/sla` runtime.
-
-### Added — payment event coverage
-
-Expanded `PAYMENT_EVENT_TYPE` catalogue to cover auth/capture and dispute
-lifecycles: `AUTHORIZED`, `CAPTURED`, `AUTH_VOIDED`, `DISPUTED`,
-`DISPUTE_WON`, `DISPUTE_LOST`, `SETTLED`.
-
-`PaymentRefundedPayload` and `PaymentReversedPayload` now carry
-`originalAmount` (and `reversedAmount` / `isPartial` on reversed) so
-consumers can detect partial vs full operations without a follow-up
-Payment lookup.
-
-## [0.7.0] - 2026-05-26
-
-### Added — payment domain primitives
-
-Three new subpaths shared across every payment-aware package
-(`@classytic/invoice`, `@classytic/revenue`, `@classytic/order`, future
-POS / AP / subscription packages). Each ships under a flat subpath so
-consumers tree-shake what they don't import.
-
-- **`/payment-method-kind`** — `PaymentMethodKind` universal vocabulary
-  (kind vs host-registered `methodCode`) so packages can speak `kind`
-  while hosts speak `code`.
-- **`/payment-allocation-status`** — orthogonal allocation +
-  reconciliation status enums for payment legs (`unallocated`,
-  `partial`, `allocated`, `over_allocated`, plus reconciliation states).
-- **`/payment-events`** — pure TS payload contracts for payment domain
-  events. Wire-only (no Zod, no runtime). Anchored on `Money` +
-  `PaymentMethodKind`; causation/correlation stay on `EventMeta`.
-
-No breaking changes.
-
-## [0.6.0] - 2026-05-17
-
-### Added — five new primitives for CRM / workflow domains
-
-Five new primitives covering gaps the existing surface didn't address.
-Each ships under its own flat subpath import (no breaking changes).
-
-- **`/phone`** — `PhoneNumber` value object + `parsePhone(input)` that
-  normalizes free-form input to E.164, captures the country calling code
-  as a discrete field, and reports failures via `Result<PhoneNumber,
-  PhoneError>`. Zero external deps; ITU-T E.164 prefixes inlined. Hosts
-  that want region-of-issue / line-type detection layer on top with
-  libphonenumber-js.
-
-- **`/status-history`** — Generic, immutable status-change log:
-  `StatusChangeEntry<TStatus>`, `appendStatus()`, `timeInStatus()`,
-  `lastTransitionTo()`. Captures `durationInPriorMs` at write time so
-  funnel-velocity dashboards ("avg time in Qualification") work without
-  re-walking timestamps. Pairs naturally with `/state-machine`.
-
-- **`/condition`** — Tiny JSON-serializable predicate DSL: `FieldCondition`
-  (`{ field, op, value }`) composed via `AllCondition` / `AnyCondition`
-  / `NotCondition`, plus a total `evaluate(condition, target)` and a
-  `validateCondition()` for boot-time schema checks. Powers SLA scope
-  filters, drip "if opened/clicked" branches, automation triggers, and
-  permission row-scopes — anywhere a predicate needs to round-trip
-  through a database.
-
-- **`/mixin`** — Additive type composition without subclassing:
-  `withMixin<T, K, M>()` / `getMixin()` / `hasMixin()` / `withoutMixin()`.
-  Lets a `Contact` carry per-domain payloads (`mixins.customer`,
-  `mixins.lead-fit`) on a typed `mixins` namespace without bloating the
-  base type. Plain object composition — no decorators, no global
-  hierarchy, no proxies.
-
-- **`/sla-policy`** — Higher-order SLA layered on the existing `/sla`:
-  priority matrix (`urgent` / `high` / `normal` → durations), first-
-  response vs rolling-response semantics, and working-hours window
-  (weekdays + start/end minute + holidays). `defineSLAPolicy()` validates
-  at boot; `evaluateSLAStatus()` produces `{ kind, responseBy,
-  remainingMs, breached }` from a policy + per-entity inputs. The
-  existing simple `/sla` interface is unchanged — `sla-policy` derives
-  concrete `SLA` instances from a policy + priority and delegates to it.
-
-### Changed — source-tree reorganization (no public-surface impact)
-
-The 24 primitive source files moved from a flat `src/*.ts` layout into
-six category folders so the package is self-documenting at the file
-system level:
-
-```
-src/
-  money/           money, currency, split-allocation, bank-transaction, payment-gateway
-  identity/        person, address, phone
-  scheduling/      period, cadence, sla, sla-policy
-  workflow/        state-machine, status-history, approval, hold, condition
-  events/          events, outbox
-  composition/     mixin, reference, context, brand, result
-```
-
-**Public subpath imports are unchanged** — `@classytic/primitives/money`,
-`@classytic/primitives/sla`, etc. still resolve. tsdown emits each entry
-to `dist/<name>.mjs` regardless of source nesting. Existing consumers
-upgrade with `npm install @classytic/primitives@^0.6.0` and a no-op
-diff.
-
-### Dependencies
-
-Still zero runtime deps. Node 22+.
-
-## [0.5.0] - 2026-05-05
-
-### Added — `/bank-transaction` subpath
-
-Canonical bank-feed shapes shared across `@classytic/fin-io` (parsers
-produce these), `@classytic/revenue` (`transactionRepository.import()`
-consumes them), and `@classytic/ledger` (`bankStatementMapper`
-consumes them). Single source of truth for the cross-package
-vocabulary; eliminates the structural mirrors fin-io and revenue
-were carrying.
-
-Exports: `BankTransaction`, `BankCounterparty`, `BankAccount`,
-`BankStatement`, `BankStatementSource`, `BankImportReport`,
-`BankImportRowError`. All anchored on `Money` from `/money` (integer
-minor units, `number`).
-
-### Added — `/payment-gateway` subpath
-
-Canonical payment-gateway data shapes. **Lets any provider package
-(Stripe, Razorpay, SSLCommerz, PayPal, bKash, Nagad, manual, …)
-implement against primitives' types without depending on revenue's
-heavyweight runtime.** Provider authors peer-dep on
-`@classytic/primitives` only.
-
-Exports: `CreateIntentParams`, `PaymentIntent`, `PaymentResult`,
-`RefundResult`, `WebhookEvent`, `ProviderCapabilities`. All `amount`
-fields are `Money` from `/money` (no flat `amount: number, currency:
-string` pairs — currency travels inside Money).
-
-The `PaymentProvider` abstract class — the *contract* every provider
-must satisfy — stays in `@classytic/revenue`. Revenue is the data
-engine that consumes providers; provider packages live elsewhere.
-
-### Architectural intent
-
-Revenue 3.0 + primitives 0.5 + fin-io 0.3 establish a clean
-three-tier seam:
-
-  1. **primitives** — pure data shapes (no runtime, no deps). Owns
-     `Money`, `BankTransaction`, `PaymentIntent`, every cross-
-     package vocabulary type.
-  2. **revenue** — the engine. Owns the abstract contracts
-     (`PaymentProvider`, `BankFeedProvider`), state machines, repos.
-     Consumes primitives shapes.
-  3. **provider packages** (`revenue-stripe`, `revenue-razorpay`,
-     `@classytic/fin-io/adapters/revenue`, …) — implementations.
-     Peer-dep on primitives. Optional peer-dep on revenue for the
-     abstract base class.
-
-This eliminates the cycle where every provider package transitively
-pulled mongoose + mongokit just to know what a `PaymentIntent`
-looked like. Tree-shaking, build times, and dep-graph honesty all
-improve.
-
-### Migration
-
-Consumers MUST import from primitives subpaths. No re-exports from
-revenue or fin-io (PACKAGE_RULES P2):
-
-```ts
-// Before
-import type { PaymentIntent, CreateIntentParams } from '@classytic/revenue';
-import type { CanonicalTransaction, Money } from '@classytic/fin-io';
-
-// After
-import type {
-  PaymentIntent,
-  CreateIntentParams,
-} from '@classytic/primitives/payment-gateway';
-import type { BankTransaction } from '@classytic/primitives/bank-transaction';
-import type { Money } from '@classytic/primitives/money';
-```
-
-Provider helper classes (`PaymentIntent`, `PaymentResult`,
-`RefundResult`, `WebhookEvent`) are deleted from revenue — they were
-trivial wrappers around plain objects. Provider implementations now
-return plain object literals matching the interface.
-
-`CreateIntentParams.amount` is now `Money` (`{ amount: number,
-currency: string }`), not the flat `amount + currency` pair. Revenue's
-repository wraps internally:
-
-```ts
-const intent = await provider.createIntent({
-  amount: { amount: params.amount, currency },
-  metadata: params.metadata,
-  ...params.paymentData,
-});
-```
-
-## [0.3.1] - 2026-05-02
-
-### Added — state-machine ↔ mongokit/sqlitekit integration
-
-State-machine primitives now compose directly with the kit-level CAS surface shipped in mongokit 3.13 and the sqlitekit follow-up. Three additions:
-
-- **`StateMachine.validTargets(from)`** — every status the aggregate can move TO from `from`. Useful for UI dropdowns ("which actions are legal right now?") and for fanning a single source state into a multi-source CAS via a repo's `claim({ from: machine.validTargets(current), to })`.
-- **`StateMachine.validSources(to)`** — every status that can transition INTO `to` (reverse adjacency). Pair with a repo's multi-source `claim({ from: machine.validSources(to), to })` when the target is fixed but the caller doesn't yet know the current state — the canonical "cancel from any non-terminal" / "error from any in-flight" pattern. Pre-computed at definition time, O(1) lookup, frozen arrays safe to expose by reference.
-- **`assertAndClaim(machine, repo, id, args)`** — pairs `machine.assertTransition` (sync domain check) with `repo.claim()` (atomic CAS) in one call. The canonical state-machine-backed CAS pattern. Skipping either layer leaves a hole: skip `assertTransition` and bad transitions reach storage; skip `claim` and concurrent writers race.
-
-```ts
-import { defineStateMachine, assertAndClaim } from '@classytic/primitives/state-machine';
-
-const ORDER_MACHINE = defineStateMachine<OrderStatus>({
-  name: 'Order',
-  transitions: {
-    draft:     ['approved', 'cancelled'],
-    approved:  ['shipped', 'cancelled'],
-    shipped:   [],
-    cancelled: [],
-  },
-});
-
-// Single-source:
-const updated = await assertAndClaim(ORDER_MACHINE, orderRepo, orderId, {
-  from: 'draft',
-  to: 'approved',
-  patch: { approvedAt: new Date() },
-  options: { organizationId: ctx.organizationId, session },
-});
-
-// Multi-source via reverse adjacency — "cancel from any legal predecessor":
-await assertAndClaim(ORDER_MACHINE, orderRepo, orderId, {
-  from: ORDER_MACHINE.validSources('cancelled'), // ['draft', 'approved']
-  to: 'cancelled',
-  patch: { cancelledAt: new Date() },
-  options: { organizationId: ctx.organizationId },
-});
-```
-
-#### `ClaimableRepo<TDoc>` — structural dependency on the kit
-
-`assertAndClaim` accepts any object matching `ClaimableRepo<TDoc>` — a structural type that mirrors the mongokit 3.13 `Repository.claim()` signature. Primitives doesn't take a hard peer dep on mongokit; consumers wire the kit they want. Sqlitekit's `claim()` matches the same shape (verified by mongokit/sqlitekit cross-conformance).
-
-```ts
-export interface ClaimableRepo<TDoc> {
-  claim(
-    id: string,
-    transition: {
-      field?: string;
-      from: unknown | readonly unknown[];
-      to: unknown;
-      where?: Record<string, unknown>;
-    },
-    patch?: Record<string, unknown>,
-    options?: Record<string, unknown>,
-  ): Promise<TDoc | null>;
-}
-```
-
-#### Why a 0.3.1 (not 0.4.0)
-
-Purely additive — no removals, no signature changes. `defineStateMachine` callers from 0.3.0 see two new methods on the returned `StateMachine` (`validTargets`, `validSources`) and one new free function (`assertAndClaim`). Backwards-compatible.
-
-## [0.3.0] - 2026-04-29
-
-### Removed — `/errors` subpath (BREAKING)
-
-- Deleted `@classytic/primitives/errors` (`ErrorContract`, `ErrorDetail`, `ErrorCode`, `ERROR_CODES`).
-- Errors are infrastructure-shaped (HTTP-coupled, repository-coupled, RFC 7807-shaped). Moved to `@classytic/repo-core/errors` where they sit next to `pagination`, `tenant`, `cache`, `hooks`, `schema` — every other repository contract — and join the existing `HttpError` throwable contract there. Single canonical home for both the wire (`ErrorContract`) and throwable (`HttpError`) shapes.
-
-#### Migration
-
-```diff
-- import type { ErrorContract, ErrorDetail, ErrorCode } from '@classytic/primitives/errors';
-- import { ERROR_CODES } from '@classytic/primitives/errors';
-+ import type { ErrorContract, ErrorDetail, ErrorCode } from '@classytic/repo-core/errors';
-+ import { ERROR_CODES } from '@classytic/repo-core/errors';
-```
-
-Field shapes and code values are unchanged — pure relocation.
-
-### Removed — `/tenant` subpath (BREAKING)
-
-- Deleted `@classytic/primitives/tenant` (`TenantConfig`, `TenantStrategy`, `TenantFieldType`, `DEFAULT_TENANT_CONFIG`, `resolveTenantConfig`, `ResolvedTenantConfig`).
-- Tenant scope is **infrastructure-shaped** (describes how queries get scoped to a database / repository), not a domain primitive like Money or Address. Moved to `@classytic/repo-core/tenant` where it sits next to `context`, `filter`, `hooks`, `schema`, `cache` — every other repository contract.
-- Mongokit and sqlitekit already peer-dep `@classytic/repo-core`, so consuming `TenantConfig` from there costs zero extra peer deps for those kits.
-
-#### Migration
-
-```diff
-- import type { TenantConfig } from '@classytic/primitives/tenant';
-- import { resolveTenantConfig, DEFAULT_TENANT_CONFIG } from '@classytic/primitives/tenant';
-+ import type { TenantConfig } from '@classytic/repo-core/tenant';
-+ import { resolveTenantConfig, DEFAULT_TENANT_CONFIG } from '@classytic/repo-core/tenant';
-```
-
-Field shapes, runtime semantics, and defaults are unchanged. The custom-strategy escape hatch (`strategy: 'custom'` + `resolve: (ctx) => filterShape`) works identically.
-
-### Removed — `/pagination` subpath (BREAKING)
-
-- Deleted `@classytic/primitives/pagination` (`PageParams`, `KeysetParams`, `SortSpec`, `SortDirection`, `OffsetPage<T>`, `KeysetPage<T>`, `AggregatePage<T>`, `emptyOffsetPage`, `emptyKeysetPage`).
-- Pagination shapes are owned by **`@classytic/repo-core/pagination`** — that package ships both the types AND the cursor encoding/skip-math/keyset-validation algorithms. Primitives' duplicate type-only declarations existed before the contract was canonised in repo-core; keeping them invited drift.
-
-### Migration
-
-```diff
-- import type { OffsetPage, KeysetPage, AggregatePage, SortSpec } from '@classytic/primitives/pagination';
-+ import type {
-+   OffsetPaginationResult,
-+   KeysetPaginationResult,
-+   AggregatePaginationResult,
-+   SortSpec,
-+ } from '@classytic/repo-core/pagination';
-```
-
-Note: shape names also changed (`OffsetPage` → `OffsetPaginationResult`, etc.) to match repo-core's naming. Field names align with what mongokit, sqlitekit, and arc already produced — those packages always returned repo-core shapes; primitives' types were a parallel, never-quite-aligned vocabulary.
-
-If you also need HTTP wire envelopes (`{ success: true } & Result`), repo-core 0.3.0 ships `OffsetPaginationResponse`, `KeysetPaginationResponse`, `AggregatePaginationResponse`, `BareListResponse`, `PaginatedResponse`, plus a `toCanonicalList()` runtime normalizer.
-
-### Why now (and why this is the only breaking change in this release)
-
-Two layers of pagination shapes drifted across the org: primitives declared one set, repo-core declared a different set, mongokit + sqlitekit + arc all consumed repo-core's. Removing primitives' set is purely subtractive — nobody internally imports from `@classytic/primitives/pagination`, and external consumers get a clear migration path to repo-core. No new ground to break later.
-
-## [0.2.0]
-
-### Added
-
-- **`/state-machine` subpath** — `defineStateMachine<TStatus>({ name, transitions, errorFactory? })`
-  declarative state-transition primitive. Replaces hand-rolled
-  `if (status === 'X') throw` blocks across kernels:
-
-  ```typescript
-  import { defineStateMachine } from '@classytic/primitives/state-machine';
-
-  const ORDER_MACHINE = defineStateMachine<OrderStatus>({
-    name: 'Order',
-    transitions: {
-      draft: ['approved', 'cancelled'],
-      approved: ['shipped', 'cancelled'],
-      shipped: [],
-      cancelled: [],
-    },
-  });
-
-  ORDER_MACHINE.assertTransition('order-1', 'draft', 'approved'); // OK
-  ORDER_MACHINE.assertTransition('order-1', 'shipped', 'draft');  // throws
-  ORDER_MACHINE.canTransition('draft', 'shipped');                // false
-  ORDER_MACHINE.isTerminal('shipped');                            // true
-  ```
-
-  Generic over the status string-union — `Record<TStatus, readonly TStatus[]>`
-  on `transitions` enforces exhaustiveness at the call site. Kernels with their
-  own typed transition error class (e.g. flow's `InvalidTransitionError`) wire
-  it via `errorFactory` so the thrown type stays domain-consistent. Standalone
-  callers get the bundled `IllegalTransitionError` (carries `entityType` /
-  `entityId` / `from` / `to` + `code: 'illegal_transition'` + `status: 422`).
-
-  First consumer: `@classytic/flow` `PROCUREMENT_MACHINE` (procurement-order
-  approve / cancel transitions).
-
-## [0.1.1]
-
-### Added
-
-- **`MinorUnits<C>` branded type on `/money`** — wire-format monetary amount
-  (integer minor units, branded with phantom currency) for HTTP/JSON fields
-  where a full `Money` struct is overkill. Erases at runtime; `MinorUnits<'BDT'>`
-  vs `MinorUnits<'USD'>` collide at compile time. Pair with structured `Money`
-  for in-process arithmetic. Constructor: `MinorUnits(amount, currency?)`.
-
-### Changed
-
-- **`/tenant` `contextKey` cascade default** — `resolveTenantConfig({ tenantField: 'branchId' })`
-  now defaults `contextKey` to `'branchId'` instead of `'organizationId'`.
-  When a host renames `tenantField`, their `OperationContext` carries the id
-  under the same key in the overwhelming majority of cases; mirroring the
-  rename is the least-surprise default. Explicit `contextKey` still overrides
-  (set when the doc field and context key genuinely diverge). Existing call
-  sites that supplied both `tenantField` and `contextKey` are unaffected.
-
-## [0.1.0]
-
-Initial release. Framework-agnostic TypeScript primitives shared across
-Classytic packages. Zero runtime dependencies. Node 22+. ESM only.
-260 tests (unit + integration) across 20 files.
-
-All subpaths are importable independently; there is **no root barrel
-re-export** — consumers must use subpath imports (`@classytic/primitives/money`,
-`/events`, `/tenant`, …) to keep the Node ESM loader + TypeScript compiler
-from walking the full module graph on any convenience import.
-
-### Subpaths
-
-- **`/money`** — `Money` type + integer-minor-unit arithmetic (`money`,
-  `addMoney`, `subtractMoney`, `multiplyMoney`, `sumMoney`, `negateMoney`,
-  `absMoney`, `compareMoney`, `equalsMoney`, `isZeroMoney`, `isPositiveMoney`,
-  `isNegativeMoney`, `isMoney`, `fromMajor`, `toMajor`); `CurrencyMismatchError`.
-- **`/currency`** — `CurrencyCode` brand; `CURRENCIES` table; `MINOR_UNIT_FACTOR`,
-  `minorUnitFactor`, `isCurrencyCode`, `toCurrencyCode`.
-- **`/address`** — `Address`, `ContactAddress`, `GeoPoint`, `GeoJsonPoint`;
-  `toGeoJsonPoint`, `fromGeoJsonPoint`.
-- **`/period`** — `DateRange`, `Period`; `isDateRange`, `isWithin`,
-  `rangeDurationMs`.
-- **`/pagination`** — `PageParams`, `OffsetPage<T>`, `KeysetPage<T>`,
-  `AggregatePage`, `KeysetParams`, `SortSpec`, `SortDirection`;
-  `emptyOffsetPage`, `emptyKeysetPage`.
-- **`/reference`** — `ExternalRef`, `DocumentRef`, `ObjectIdLike`, `IdLike`;
-  `idToString`, `isExternalRef`.
-- **`/context`** — `OperationContext` (the identity + tracing bag every
-  Classytic package accepts); `ActorRef`.
-- **`/events`** — `EventMeta` (arc v2.9-aligned — 13 fields including
-  `schemaVersion`, `causationId`, `partitionKey`, `source`, `idempotencyKey`,
-  `aggregate`), `DomainEvent<T>`, `EventTransport`, `EventHandler`,
-  `EventLogger`, `DeadLetteredEvent<T>`, `PublishManyResult`; `createEvent`,
-  `createChildEvent` (auto-chains `causationId`, inherits `correlationId` /
-  `userId` / `organizationId` / `source` / `idempotencyKey` — NOT
-  `aggregate`), `matchEventPattern`.
-- **`/outbox`** — transactional outbox **contract** mirrored from arc v2.9:
-  `OutboxStore`, `OutboxWriteOptions`, `OutboxClaimOptions`,
-  `OutboxAcknowledgeOptions`, `OutboxFailOptions`, `OutboxErrorInfo`,
-  `OutboxFailureContext`, `OutboxFailureDecision`, `OutboxFailurePolicy`;
-  `OutboxOwnershipError`, `InvalidOutboxEventError`. **Contract only** —
-  runtime (`EventOutbox`, `MemoryOutboxStore`, `MongoOutboxStore`,
-  `exponentialBackoff`) ships in `@classytic/arc/events` and
-  `@classytic/arc/events/mongo`. Domain packages implement `OutboxStore`
-  without peer-depping arc.
-- **`/tenant`** — `TenantConfig` (with `strategy: 'field' | 'none' | 'custom'`
-  and optional `resolve(ctx)` for custom strategy), `TenantFieldType`,
-  `TenantStrategy`, `ResolvedTenantConfig`; `DEFAULT_TENANT_CONFIG`,
-  `resolveTenantConfig`. Field names match `@classytic/mongokit`'s
-  `MultiTenantOptions` so a resolved config threads into
-  `multiTenantPlugin({ ... })` without translation.
-- **`/result`** — `Result<T, E>`; `ok`, `err`, `isOk`, `isErr`, `unwrap`,
-  `mapResult`, `mapError`.
-- **`/brand`** — `Brand<T, B>`, `Prettify`, `DeepPartial`, `DeepReadonly`,
-  `RequireKeys`, `OptionalKeys`, `NonEmptyArray`.
-- **`/errors`** — `ErrorContract`, `ErrorDetail`, `ErrorCode`, `ERROR_CODES`.
-- **`/split-allocation`** — `allocate(total, subjects, method)` returns a
-  `SplitResult` whose parts sum exactly to `total` via the largest-remainder
-  method. Methods: `equal`, `by-qty`, `by-weight`, `by-volume`, `by-value`,
-  `by-percent`. Deterministic — identical inputs produce identical outputs;
-  ties broken by input order. Handles negative totals by sign-flipping.
-  Exports: `allocate`, `isBalanced`, `SplitAllocationError`, types
-  `SplitMethod`, `SplitSubject`, `SplitPart`, `SplitResult`,
-  `SplitAllocationErrorCode`.
-- **`/approval`** — `ApprovalChain` value object with pure FSM helpers.
-  Supports sequential and parallel chains, per-step quorum
-  (`requiredApprovals`), threshold-gated skips. Any step-level rejection
-  rejects the chain; all steps `approved`|`skipped` approves it. Exports:
-  `createChain`, `applyDecision`, `skipStep`, `nextPendingStep`,
-  `pendingSteps`, `isApproved`, `isRejected`, `isPending`, `decisionCount`,
-  `ApprovalError`, types `ApprovalChain`, `ApprovalStep`, `ApprovalDecision`,
-  `Approver`, `ApprovalStatus`, `ChainOrder`, `ApprovalThreshold`,
-  `CreateChainInput`, `CreateStepInput`, `DecisionInput`, `ApprovalErrorCode`.
-- **`/cadence`** — recurrence spec with pure UTC date arithmetic. Kinds:
-  `daily`, `weekly`, `monthly`, `yearly`, `cron` (opaque — host parses).
-  `dayOfMonth: 31` snaps to the last day of shorter months; Feb 29 snaps
-  to Feb 28 in non-leap years. Exports: `nextOccurrence`,
-  `occurrencesBetween`, `validateCadence`, `CadenceError`, types `Cadence`,
-  `DailyCadence`, `WeeklyCadence`, `MonthlyCadence`, `YearlyCadence`,
-  `CronCadence`, `CadenceKind`, `IsoWeekday`, `CadenceErrorCode`.
-- **`/hold`** — `HoldReason` + collection helpers. Host embeds
-  `HoldReason[]` on documents. Exports: `addHold`, `resolveHold`,
-  `activeHolds`, `resolvedHolds`, `isOnHold`, `hasActiveHoldOfCode`,
-  `HoldError`, types `HoldReason`, `HoldActor`, `HoldActorKind`,
-  `AddHoldInput`, `ResolveHoldInput`, `HoldErrorCode`.
-- **`/sla`** — target duration + breach-detection arithmetic. Policy is
-  carried but not executed — host reads `breachPolicy` (`warn` | `escalate`
-  | `block`) and acts. Exports: `breachedAt`, `remainingMs`, `elapsedMs`,
-  `isBreached`, `consumedFraction`, `validateSLA`, `SLAError`, types
-  `SLA`, `BreachPolicy`, `SLAErrorCode`.
-- **`/person`** — person/contact shapes shared across order, crm, hr.
-  `PersonName`, `ContactInfo`, `EmergencyContact`, `Gender`; formatters
-  `formatFullName`, `formatDisplayName`.
-
-### Design principles
-
-- Pure functions on plain data. No I/O, no Mongo, no events.
-- Immutable return values. Helpers return new collections; originals are never mutated.
-- Deterministic. Same inputs → same outputs. Clock sources injected
-  explicitly so tests don't flake.
-- Errors are typed. Every module ships an `<Name>Error` class with a
-  discriminated `code: <Name>ErrorCode` string union.
-- Subpath-imports only — no root barrel.
-
-### Ownership direction
-
-Arc is the source of truth for event types. Primitives **mirrors** arc
-(`@classytic/arc/events/EventTransport.ts`, `@classytic/arc/events/outbox.ts`)
-verbatim for `EventMeta`, `DomainEvent`, `EventTransport`,
-`DeadLetteredEvent`, `EventLogger`, `PublishManyResult`, and the full
-`OutboxStore` contract surface. Helpers (`createEvent`, `createChildEvent`)
-are allowed to be more defensive but match arc's signatures.
-
-Packages that peer-dep `@classytic/primitives` can write
-`class FooOutboxRepository implements OutboxStore` without pulling arc into
-their dependency graph; hosts import `EventOutbox` + `MongoOutboxStore` from
-arc and wire them against package-provided stores.
+# Changelog
+
+Format based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
+adhering to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
+
+## 0.30.0 — 2026-10-11
+
+### Added — `./durable-events`: the declaration and the ONE kernel dispatch
+- `EventDurability`, `DurableEventSpec`, `FactExpectation`; `eventIdFor` / `durableEventMeta` (`${name}:v${version}:${occurrence}`); `requireDurableContext` + `NonDurablePublishError`; `UnmanagedSessionError`.
+- `dispatchDomainEvent(channels, unit, event, durability)` saves the outbox row in the session, defers to commit, and acknowledges after an inline publish.
+- `flushDeferredEvents` delivers a committed queue and acknowledges every delivered row in ONE `acknowledgeMany`.
+Pinned by `tests/unit/durable-events.test.ts`.
+
+### Added — `./outbox`: `OutboxConflictError`, `outboxPayloadFingerprint`, optional `OutboxStore.acknowledgeMany`
+
+### ⚠ Changed
+- `createChildEvent` no longer inherits the parent's `idempotencyKey`. A child is its own occurrence; an inherited key made a store drop the second of two child types.
+- `MemoryOutboxStore.save`: the same id with the same payload is a no-op, and a different payload throws `OutboxConflictError`.
+
+## 0.29.0 — 2026-10-08
+
+### Added — `./signed-token`: ES256 compact JWS over WebCrypto
+`signToken` / `verifyToken` for a server and an offline device, no dependency. `typ` is required on both
+sides and a token without `exp` is refused; every failure is a named refusal (`bad_signature`,
+`wrong_type`, `unknown_key`, `expired`, …), never a throw. Signing keys import non-extractable.
+Pinned by `tests/unit/signed-token.test.ts` (the expiry rule is falsified).
+
+### Added — `itemStockOf` (`./item-stock`): one item's stock, the one summary every reader shows
+`itemStockOf(item, rowOf, { defaultReorderPoint })` builds `{ quantity, inStock, lowStock, valueMinor?,
+variants? }` from rows filed under stock keys (`stockKeyFor`). A row reached by both the item-level
+and a variant lookup counts once; production's sole variant keys by the product id. It is pure and
+browser-safe, so the server's catalog bridge and an offline till compute the same answer. Pinned by
+`tests/unit/item-stock.test.ts`; collapsing the row de-duplication fails three of its cases.
+
+### Added — maker-checker on `ApprovalChain` (`./approval`)
+`ApprovalChain.request` (`{ by, at, basis? }`) records who asked and what the approval covers.
+`attachRequest(chain, request)` removes the requester from every step's approvers and throws
+`NO_INDEPENDENT_APPROVER` when a step is left below its quorum. `applyDecision` throws
+`SELF_APPROVAL` when the requester approves (rejecting their own request — a withdrawal — is
+allowed). `applyDecision` and `skipStep` now preserve `request` and `notRequiredReason` instead of
+rebuilding the chain from `{ order, steps, status }`.
+
+### Added — `diffEventSchema` (`./events`)
+Compares two versions of an event payload's JSON Schema and marks each change breaking or not
+for an existing CONSUMER (forward compatibility): a required field removed, a field made optional,
+a widened type or enum break; additions and narrowing do not. Payloads are open for extension, so
+`additionalProperties: false` is disregarded. Nested objects, array items and record values recurse.
+
+### Added — `createEventRouter` (`./events`)
+An index over event subscriptions with `matchEventPattern`'s exact semantics — same values, same
+order, deduped — in amortized O(1) per route: candidates are only the name's separator prefixes,
+and results are memoized per name (bounded) until `add`/`remove`. Pinned by a property test
+against the scan. Transports and graph analyses route through it instead of scanning.
+
+### Added — `readEnv`: strict configuration reads (`./environment`)
+`readEnv(process.env)` → `bool | int | number | oneOf | list`, plus `InvalidEnvError`. Unset or
+blank → the declared default; a value that is SET but unrecognised throws, naming the variable
+and what it accepts. Booleans take `true/false`, `1/0`, `yes/no`, `on/off`; numbers take bounds.
+A hand-rolled `=== 'true'` or `parseInt` widens a typo to the default and says nothing.
+
+### Added — placeholder phone, one definition (`./phone`)
+`PLACEHOLDER_PHONE_PREFIX`, `placeholderPhone(id)`, `isPlaceholderPhone(v)`, `shownPhone(v)`.
+A record whose `phone` is required and unique but unknown stores `pending_<id>`; the writer
+(`@spinekit/order`) and every reader (be-prod party linking, the SDK, commerce-ui, pos-ui,
+the storefront) now share this instead of five hand-written `startsWith('pending_')`.
+
+### Deprecated: `@classytic/primitives/idempotency` — moved to `@classytic/repo-core/idempotency`
+
+The claim contract had no consumers here. It now lives beside `repo-core/lock` with its store
+port, `runIdempotent`, the reference store and the conformance suite, so every persistence kit
+implements it without depending on this package. Removed in the next minor.
+
+### Changed: `canonicalJson` refuses nesting deeper than 100 levels
+
+`MAX_CANONICAL_DEPTH` (MongoDB's document limit). Deeper input throws `CanonicalizeError`
+instead of recursing — a hostile input used to overflow the stack.
+
+### Added: `@classytic/primitives/line-settlement` (incl. `shareOfSettled`)
+
+One definition of "how much of this document line is still open". A line's
+quantity splits into kernel-named outcome buckets (fulfilled / returned /
+received / written off …); `outstandingOf`, `settlementState`,
+`documentSettlementState` and `settledEntirelyIn` derive the rest, and
+`planSettlement` REFUSES a non-integer or over-outstanding quantity instead of
+trimming it (`clampSettlement` exists for surfaces whose contract is "take what
+fits"). Order, flow, purchase and transfer each derived this by hand; transfer
+adopts it first.
+
+### Added: `OutboxClaimedEvent.attempts`, the durable attempt count
+
+`OutboxFailureContext.attempts` was documented as "tracked in-process by the
+relay; resets on restart", and that is exactly where a failure policy is wrong.
+`attempts >= 5 → deadLetter` never fired for a poison event that crashes the
+relay (it came back at 1 on every boot), and across N relays sharing a store it
+meant 5×N deliveries. Stores already increment a per-row counter atomically
+inside the claim; they had no way to report it.
+
+`OutboxClaimedEvent` (the fenced-claim result) gains an optional
+`attempts?: number`, which counts every claim including the one that returned it.
+`MemoryOutboxStore` reports it. Optional and additive: stores that don't count
+leave it absent, and relays fall back to their own count and take the larger of
+the two.
+
+## 0.28.0
+
+### Added: a `pass` monetization kind (paid once, grants a term, never renews)
+
+The union could say "charged once" (`one_time`, a single price, no span) and
+"recurring" (`subscription`, priced spans that renew). It could not say
+"prepaid term": a course access pass, a gym or transit pass, a parking permit,
+a term licence. Hosts selling one reached for `subscription`, the only shape
+that carried priced spans, and then had to explain at every read that it does
+not actually renew. `toRevenueMonetizationType` would also have settled those
+sales as recurring income, forecasting revenue nothing was scheduled to bill.
+
+`PassMonetization { type: 'pass', passes: AccessPass[] }`, where an
+`AccessPass` is a key, a label, a price, an optional "was" price, and a
+`duration` + `durationUnit`. Shaped like `SubscriptionPlan` so a term means
+the same thing everywhere, kept separate because a plan carries `trialDays`
+and a renewal discount that a pass has none of.
+
+Also: `isPassMonetization`, `defaultPassOf` (the cheapest term, ties broken on
+the shorter SPAN so the headline is deterministic), and `approxTermDays` for
+ordering and labels only. `unitPriceOf` returns the cheapest term's price for
+a pass, `compareAtPriceOf` now returns the "was" of that SAME term so a card
+can never strike one term's price through beside another's, and
+`toRevenueMonetizationType('pass')` is `purchase`.
+
+A perpetual sale stays `one_time`: "forever" is the absence of a term, not a
+very long one.
+
+### Added — `CashDestination` on `./money`
+
+`CASH_DESTINATIONS` (`'cash' | 'petty_cash'`), `isCashDestination` and
+`cashDestinationLabel`: WHERE received cash landed, as an intent — never a
+chart-of-accounts code. One vocabulary for a request body, a domain event, a
+client SDK and a settlement dialog, so the accounting tier alone maps it to the
+active chart and a renumber cannot reach any of them. Added because a client
+carried a literal bank code the chart had already renumbered, and every
+settlement it sent was refused.
+
+## 0.27.0
+
+### Added — `./csv` subpath: one CSV cell escaper for the whole fleet
+
+`escapeCsvCell(value)` and `neutralizeCsvFormula(value)`. Six CSV producers
+across ledger, the spine and the hosts each hand-rolled RFC 4180 quoting, and
+not one neutralized a leading `= + - @ \t \r` — so a customer-supplied payment
+reference or party name reached the accountant's spreadsheet as a live formula.
+Quoting does not help: a spreadsheet strips the quotes and evaluates what is
+left, which is why the two defences ship as one module.
+
+The two functions are separable on purpose. A consumer holding a real CSV
+writer (`csv-stringify`) takes `neutralizeCsvFormula` only, so its quoting is
+not done twice; a hand-rolled `rows.map(...).join(',')` takes `escapeCsvCell`.
+
+**Well-formed numbers are exempt.** Prefixing `-1500` would make every negative
+amount text and break `SUM` in every finance export — a fix that quietly
+corrupts the product is worse than the hole. The exemption requires the WHOLE
+cell to be a number, so `-2+3+cmd|'/c calc'!A0` is still neutralized.
+
+Not for a machine-to-machine upload contract with a fixed column spec (a
+courier's bulk-import CSV): there the receiver is a parser, not a spreadsheet,
+and an added `'` corrupts the payload.
+
+## 0.26.4
+
+### Added — `./price-basis` subpath: measured pricing across physical dimensions
+
+New `PriceBasis` type (`FixedPriceBasis | MeasuredPriceBasis`) models how a quoted price spans a sellable quantity — flat fixed or per-unit-of-measure (mass, duration, volume, length, area). Factory functions `fixedPriceBasis()` and `measuredPriceBasis(dimension, qtyUnit, priceUnit, qtyPerPriceUnit)` construct the two shapes; `isPriceBasis()` is a structural guard for DB/wire boundaries. `resolvePriceBasisTotal(unitPriceMinor, quantity, basis?)` converts a minor-unit price × integer quantity to a final minor-unit total using half-even rounding (`roundedDiv` internally) and throws `PriceBasisError` on unsafe inputs or overflow.
+
+### Added — `./unit-cost-rate` re-exports `RoundingMode`
+
+`RoundingMode` (`'half-even' | 'half-up'`) is now defined in the new internal `rounded-div.ts` and re-exported from `unit-cost-rate`. No breaking change — existing consumers who import `RoundingMode` from `@classytic/primitives/unit-cost-rate` continue to work.
+
+### Changed — `validateAddress` / `isAddressValid` accept `Address` directly
+
+Both functions now accept `Address | PartialAddress | null | undefined`, removing the need for a cast when validating a fully-typed value object.
+
+## 0.26.3
+
+### Added — `PaymentProviderPort.wantsRawBody`
+
+Optional `readonly wantsRawBody?: boolean` on the payment gateway port.
+Providers that verify an HMAC signature (Stripe signs the bytes it transmitted)
+must declare `true` — the transport then hands the raw body rather than the
+parsed object. Absent / `false` preserves the historical behaviour (parsed
+body), which is correct for manual, cash and terminal providers that sign
+nothing.
+
+## 0.26.2
+
+### Added — `stockKeyFor` (`/item-identity`): one stock-key precedence for every package
+
+The product-level stock key, with a single precedence:
+
+1. the variant's stamped `skuRef` — assigned once at birth, immutable
+2. a **sole** variant with no stamp → the product's `_id`, never the label
+   (legacy writers derived `variantSku || productId`, and for one variant that
+   key IS the `_id`)
+3. a **sibling** with no stamp → its label, which genuinely is its legacy key
+4. nothing addressable → the product's `_id`
+
+The caller declares only what it HAS — `{ kind: 'sole' | 'label' | 'variant' }`
+— so the sole-vs-sibling decision is internal and a caller cannot make it wrong.
+That choice previously lived at every call site across three packages, in two
+functions with different fallbacks; they disagreed on a sole, unstamped,
+labelled variant, and the read then looked up a key no quant is stored under.
+
+It lives here, below every domain, because `@spinekit/order` and
+`@spinekit/catalog` must not import a peer domain — the rule sitting one layer
+too high is exactly why each package re-derived it.
+
+## 0.26.1
+
+- `./testing/stock-key-seam` — new subpath; `scanStockKeySeam`, `assertStockKeySeam`, `takesStockKeyFromLabel`, `STOCK_KEY_SEAM_FIXTURES` for conformance-scanning packages that resolve stock keys
+
+## [0.26.0] — 2026-08-27
+
+### Added
+
+- **`./deadline` subpath** — `Deadline` value type + helpers for date-bounded
+  obligations (reservations, SLA windows, contract terms).
+- **`./stock-level` subpath** — canonical stock-level representation and helpers.
+- **`./item-facts` subpath** — cross-kernel item fact contract (weight, dimensions,
+  hazmat flags) that catalog declares and logistics/inventory reads.
+- **`StockKey` branded type** (`./item-identity`) — prevents a bare merchandising
+  label or ObjectId being passed where a quant key is expected; the only source is
+  `stockKeyOf()` / `requireStockKey()`. Return types of both functions updated.
+- **Upfront collection policy** (`./monetization`) — `UpfrontCollectionPolicy`
+  (`'full_payment' | 'deposit' | 'deferred'`) + `UPFRONT_COLLECTION_POLICIES`
+  const. Cross-kernel: catalog declares the policy, order turns it into tenders,
+  revenue settles them.
+
+## [0.25.0] - 2026-08-24
+
+### Added
+
+- **`@classytic/primitives/item-identity` subpath** — `SkuFields`, `stockKeyOf`, `skuDisplay`. `sku` is a label; `skuRef` is the stock identity — the two fall back in opposite directions and getting one backwards type-checks while returning a plausible wrong string. `stockKeyOf` resolves `skuRef ?? sku ?? fallback` (the key to reserve, decrement and post against); `skuDisplay` resolves `skuLabel ?? sku ?? skuRef` (what a human reads). Canonicalises the per-package `variantSku || String(productId)` derivations into one place.
+
+## [0.24.0] - 2026-08-24
+
+### Added — `@classytic/primitives/address`: capture-format system
+
+`AddressFormat`, `AddressFieldDescriptor`, `AddressFieldKey`, `PartialAddress`, `AddressFieldError` — a description of HOW a country collects a `ContactAddress`: field labels, order, requiredness, subdivision taxonomy hooks, and postal code pattern. Country packs (e.g. `@classytic/bd-areas/address-format`) live beside their taxonomies; this subpath owns only the generic shape.
+
+`validateAddress(address, format)` — validates a partial address against a format; returns `AddressFieldError[]`. Enforces `required` on both canonical and extension fields identically, and checks `postalCodePattern` only when a value is given and a pattern is declared. `isAddressValid` and `describeAddressError` are convenience wrappers.
+
+Every `key` in a descriptor is a field of `ContactAddress`; co-location with the value object is what makes a mismatch obvious instead of invisible.
+
+## [0.23.0] - 2026-08-15
+
+### Added — `./approval`: `notRequiredChain(reason)` + `isNotRequired(chain)`
+
+The chain for a subject that needs **no** approval — zero steps, already resolved. The
+explicit, greppable answer to "the policy engine ran and decided this document does not
+require approval", which callers previously had no way to express.
+
+This states a conclusion the module already reached rather than inventing one:
+`computeChainStatus` folds an empty step list to `'approved'` — "nothing outstanding" is
+what approved MEANS here, and that fold predates this function. `createChain` still throws
+`EMPTY_STEPS`: building a ROUTING chain that routes nowhere is a different mistake and
+stays refused. The two are deliberately not interchangeable, so reaching "approved with no
+approvers" requires saying so by name.
+
+**Why a chain rather than leaving the field unset.** Absence is the one input that occurs
+by accident — a schema that never declared `approvals` strips it on write — and
+`assertApproved` must keep treating that as `CHAIN_MISSING`. A present zero-step chain is
+distinguishable from a lost chain (absent) *and* from a granted one (has steps and
+decisions), so existing gates keep working through `isApproved` alone. No consumer needs a
+second predicate, which is what would drift.
+
+`isNotRequired` tests the STRUCTURE (`steps.length === 0`), not the new
+`notRequiredReason` field: a subject schema that never declared that field drops it on
+write, so a predicate reading it would answer differently before and after the save. The
+reason is provenance for the audit trail only — pinned by a test that strips it and asserts
+the predicate is unmoved.
+
+**The hazard, stated in the docblock:** this makes `isApproved` true with no human
+decision. It is correct only where a policy engine ASKED and got "no policy applies" —
+never as a fallback for a lookup that failed or was never configured. First consumer:
+`@classytic/arc-approval` 0.4.0 `whenNoPolicy: 'not-required'`.
+
+## [0.22.0] - 2026-08-12
+
+### Added — outbox fencing tokens (`claimPendingFenced`)
+
+- **`OutboxStore.claimPendingFenced?`** (additive, feature-detected — mirrors repo-core's `LockAdapter.tryAcquireFenced`): same atomic semantics as `claimPending`, returning `OutboxClaimedEvent { event, fencingToken }`. The token is minted by the STORE per claim epoch — a takeover after lease expiry mints STRICTLY GREATER, so a stale ex-holder's `acknowledge`/`fail` (now accepting `fencingToken?`) throws `OutboxOwnershipError`. This closes the overlap that consumer-id ownership checks narrow but cannot close: holder ids recur; epochs do not.
+- `MemoryOutboxStore` implements it; `claimPending` delegates to the fenced path (one claim implementation). Stale-ack rejection is mutation-proven.
+
+### Added — `./period`: `shiftPeriod(period, n)` — calendar arithmetic without instants
+
+Shift a MONTH or QUARTER period by `n` of its own units (negative = back); YEAR shifts by years; a `custom` (ad-hoc start/end) period has no unit to shift and throws rather than guessing one. No `Date` is constructed and no zone is consulted — shifting a month is pure `(year, month)` arithmetic, and the `timezone` rides along unchanged so the shifted period still resolves against the calendar it was defined in.
+
+Exists because both inline shapes callers reach for are wrong the same way: `new Date(Date.UTC(y, m - 1 - n, 1))` answers a business-calendar question with a UTC instant (six hours off at both ends in Asia/Dhaka — the bug class this module's header documents five instances of), and hand-rolled `y * 12 + (m - 1) - n` integer math is a third copy of a rule `Period` already owns, with the year rollover exactly where copies go wrong. First consumer: be-prod `business-date.ts` ("two filing months ago").
+
+**Versioning note:** this export briefly existed in a working copy labelled 0.21.0 after 0.21.0 had already been published without it — same version, different surface. 0.22.0 restores the invariant that a version number names exactly one contract; `>=0.22.0` is the correct floor for any consumer importing `shiftPeriod`.
+
+### Docs — `./environment` docblock refinement (no API change)
+
+The 0.21.0 entry below originally listed `isProduction` / `isDevelopment` / `isTest`; the published surface is `classifyEnv`, `EnvClass`, `isProductionEnv` only — `isProductionEnv` is deliberately the single predicate (consumers needing all three buckets build their own triple from `classifyEnv`). The entry below is corrected in place; runtime and declarations are byte-identical to published 0.21.0 apart from comments.
+
+## [0.21.0] - 2026-08-10
+
+### Added — `./environment`: canonical `NODE_ENV` classification
+
+`classifyEnv(raw)` — total function mapping a raw `NODE_ENV` string to `EnvClass = 'development' | 'test' | 'production'`. Every input yields exactly one bucket; no fourth state or `undefined`. Handles both spellings per bucket (`production`/`prod`, `test`/`qa`), trims whitespace (`.env` files and CI variables routinely carry trailing newlines), and defaults unknown values to `development` — the safe-fail direction that keeps local tooling functional.
+
+Collapses four copies across the codebase that each missed at least one spelling: `be-prod classifyEnv`, `gym businessMode()`, arc's CLI config template (`z.preprocess`), and arc's four raw comparison sites. The missed spellings were all in the dangerous direction — a `NODE_ENV=prod` instance that matched nothing fell through to `development`, opening security gates (`Secure` cookie flag, arc's dev preset, non-transactional money writes, raw 500 messages to the client) in production.
+
+Exports: `classifyEnv`, `EnvClass`, `isProductionEnv`.
+
+### Added — `publint` gate on build
+
+`publint: true` in `tsdown.config.ts` — catches export-map entries pointing at non-existent dist files at build time rather than at consumer install time.
+
+## [0.20.0] - 2026-08-05
+
+### Added — `./period` OWNS timezone-aware period resolution (was two copies above the kernels)
+
+`Period` was `{ year, month?, quarter? }` with no timezone and no resolver, so it
+could not express "the business month of August in Asia/Dhaka". The resolver
+existed TWICE, both ABOVE the kernels — `@spinekit/kit/period`
+(`createBusinessCalendar`) and be-prod's `#lib/utils/business-date` — which meant
+a kernel could reach neither and re-derived the boundary inline with `Date.UTC`.
+Five silent UTC-month defects came out of that (`aggregateMonthlyVat`, the
+sales-fact reconciler, and three in bd-tax Mushak).
+
+This is a MOVE of the proven `createBusinessCalendar` logic, not a re-derivation:
+same civil-date round-trip through `/timezone`, same half-open ranges, same
+DST-exactness.
+
+- **`Period.timezone?`** — the IANA zone a period is DEFINED in, so a persisted or
+  transported period keeps its own calendar.
+- **`resolvePeriod(period, timezone?)`** → half-open `DateRange`. Throws when no
+  zone is available anywhere (never silently answers in UTC) and throws when the
+  period's zone and the caller's zone DISAGREE.
+- **`resolveDay` / `resolveMonth` / `resolveQuarter` / `resolveYear` /
+  `resolveDateSpan` / `dayStart`** — the granularity-specific forms. All return
+  HALF-OPEN `[start, end)`, which tiles exactly.
+- **`periodOf(instant, granularity, timezone)`** — the inverse: which filing
+  period an instant belongs to on the LOCAL calendar.
+- **`parsePeriod` / `formatPeriod` / `granularityOf` / `periodTimeZone`** —
+  `'YYYY'` · `'YYYY-MM'` · `'YYYY-Qn'`. `parsePeriod` REFUSES a `'YYYY-MM-DD'`
+  rather than reading it as its month (a ~30× widening).
+- **`inclusiveEnd(range)`** — `end - 1ms` for legacy `$lte` call sites, derived
+  from the resolved end so it stays DST-exact.
+- **`PeriodError`** + **`PeriodErrorCode`** (`AMBIGUOUS_PERIOD` ·
+  `INCOMPLETE_RANGE` · `INVALID_FIELD` · `MISSING_TIMEZONE` · `TIMEZONE_CONFLICT`
+  · `INVALID_LABEL`).
+
+A `Period` carrying two resolution shapes (`month` + `quarter`, or either with
+`start`/`end`) now throws instead of preferring one.
+
+### Added — `./idempotency`: the claim + lease CONTRACT
+
+Three kernels hand-rolled incompatible versions of one state machine — `cart`
+(`in_flight`/`succeeded`/`failed`, the only one with crash-window leasing),
+`catalog` offers (`pending`/`completed`, no lease), `contract` amendments
+(implicit). Types + PURE helpers only; no persistence, no collection, no driver.
+
+- **`ClaimOutcome = 'claimed' | 'replayed' | 'in_flight'`** — three-valued, and
+  `in_flight` is NOT a failure. Mapping a live concurrent lease onto an error
+  licenses a retry, which is a double-apply.
+- **`IdempotencyClaim`** (`identity` · `requestFingerprint` · `state` ·
+  `leaseToken` · `leaseExpiresAt` · `attempts` · `result`), **`ClaimResult`**,
+  **`ClaimState`**, **`IdempotencyIdentity`**.
+- **`decideClaim(existing, request)`** — the whole state machine as one pure
+  function. Throws `FINGERPRINT_MISMATCH` (same key, different body) and
+  `MISSING_REPLAY` (terminal record with no stored result) rather than answering.
+- **`completeClaim` / `renewLease`** — both refuse when the caller lost its lease,
+  so a superseded attempt cannot overwrite the winner's result.
+- **`identityKey`** (canonical-JSON composite, so a delimiter inside a scope value
+  cannot collide two identities), **`fingerprintRequest`**, **`newLeaseToken`**,
+  **`isLeaseExpired`**, **`holdsLease`**, **`isInFlight`**, **`DEFAULT_LEASE_MS`**.
+
+### Changed — BREAKING: five API decisions taken before spine wiring starts
+
+- **`./cadence` — `Cadence.timezone` is now AUTHORITATIVE, not display-only.**
+  It was documented as decorative while `stepMonthly`/`stepYearly` used
+  `getUTCMonth()`/`getUTCDate()`, so `timezone: 'Asia/Dhaka'` + `dayOfMonth: 1`
+  with a midnight anchor placed every occurrence on the local **2nd**. All four
+  kinds now run on the zone's wall clock when one is set (daily advances LOCAL
+  days — 23h/25h across a DST transition), and stay UTC when it is absent.
+  `validateCadence` rejects an unresolvable zone with the new
+  `INVALID_TIMEZONE`. **Behavioural change for any persisted cadence that
+  already carries a `timezone`** — the occurrences move to where the field
+  always claimed they were.
+- **`./approval` — `assertApproved` no longer passes on a `null` chain.** It now
+  throws the new `CHAIN_MISSING`. Absence is not approval, and a mongoose schema
+  that never declared `approvalChain` strips it on write, so the gate could not
+  tell "no approval configured" from "approval silently lost". The explicit
+  opt-out is the new **`assertApprovedIfPresent`**. The second parameter is now
+  an options object (`{ message }`), not a bare string.
+- **`./payment-events` — every payload carries an `eventType` discriminant, and
+  there is a `payment.unknown` event.** `PaymentInitiatedPayload` and
+  `PaymentSucceededPayload` were structurally IDENTICAL with no discriminant on
+  the union, so a handler typed for "funds received" accepted "money in flight".
+  Adds **`PaymentUnknownPayload`** (the port contract's third outcome had no
+  event, so an unobserved result had to be forced into SUCCEEDED or FAILED),
+  **`PaymentEventPayloadMap`**, **`isPaymentEvent`**, **`isFundsReceived`**, and
+  three compile-time drift guards over the map.
+- **`./currency` — `convertWithSnapshot` / `reverseWithSnapshot` take and return
+  `Money`.** A bare `number` meant nothing checked the amount was denominated in
+  `fx.sourceCurrency`; a USD→BDT snapshot applied to a EUR amount type-checked
+  and returned a plausible wrong number. Both now throw `CurrencyMismatchError`
+  on a mismatch, handle differing minor-unit exponents (USD 2 → JPY 0), and take
+  an explicit **`FxRounding`** (`'half-away-from-zero'` default · `'floor'` ·
+  `'ceil'`) instead of returning an unrounded float for each caller to round
+  differently.
+- **`./money` — `Money.currency` is `CurrencyCode`, not `CurrencyCode | string`.**
+  A union with `string` collapses to `string`, so the brand was decoration on
+  the one type where it matters and `money()` validated nothing. `money()`,
+  `fromMajor()` and `sumMoney()` now validate and brand; `isMoney` format-checks
+  the code instead of `typeof === 'string'`. New **`currencyCode(value)`**
+  (throwing constructor) and **`InvalidCurrencyCodeError`**.
+  **`CurrencyMismatchError` is now DECLARED in `./currency`** (so the FX seam can
+  throw it without an ESM cycle) and re-exported from `./money`, so existing
+  imports are unaffected.
+
+Migration: a `Money` object literal now needs a branded code —
+`money(1999, 'USD')`, or `{ amount: 1999, currency: currencyCode('USD') }`.
+
+### Changed — six silent-permissiveness defects now FAIL instead of answering
+
+Every item below previously returned a plausible-looking value where it should
+have errored. All are behaviour changes for malformed input only; correct input
+is unaffected. Falsifying tests live in `tests/unit/fail-loud-regressions.test.ts`.
+
+- **`./period` — an Invalid Date no longer reports "no overlap".** `rangesOverlap`,
+  `isWithin` and `rangeDurationMs` throw the new `DateRangeError` on a non-finite
+  endpoint; `isDateRange` returns `false` for one. Every comparison against `NaN`
+  is `false`, so a corrupt range used to answer "no conflict" — the PERMISSIVE
+  answer for the three documented consumers (order booking collision, flow
+  reservation-window collision, promo campaign overlap), i.e. a silent double-book.
+- **`./currency` — the ISO 4217 exponent table is now exhaustive.** All 0-decimal
+  (`BIF CLP DJF GNF ISK KMF KRW PYG RWF UGX UYI VND VUV XAF XOF XPF`), 3-decimal
+  (`BHD IQD JOD KWD LYD OMR TND`) and 4-decimal (`CLF UYW`) currencies are listed,
+  so `minorUnitFactor`'s `?? 100` fallback is reachable only by a code that is not
+  a real ISO currency. Previously `fromMajor(1500, 'GNF')` stored 150 000 minor
+  units — a 100× error that looks like a number. New `isKnownCurrency(code)`
+  distinguishes "listed" from "assumed" for boot-time config validation.
+- **`./proration` — a missing allocated part throws instead of becoming `0`.** The
+  `?? 0` fallbacks in `splitByPeriodFraction` / `allocateMoneyByFraction` are now
+  `ProrationError('ALLOCATION_INVARIANT')`. A zero credit and a genuinely
+  fully-consumed period were indistinguishable downstream.
+- **`./state-machine` — `assertAndClaim` rejects an EMPTY `from` list.**
+  `validSources(to)` returns `[]` for an unreachable status, and the documented
+  `from: machine.validSources(x)` idiom then ran the assert loop zero times and
+  handed the CAS an empty allow-list. An empty allow-list must never mean "any".
+- **`./timezone` — the offset regex is anchored.** An offset label the runtime
+  emits in an unexpected shape (e.g. an unpadded `GMT+5:30`) matched the optional
+  bare-`GMT` branch and resolved to **UTC**, silently shifting every business-day
+  boundary derived from it. Unrecognised labels now throw; `\d{1,2}` accepts the
+  unpadded variant rather than mis-reading it.
+- **`./canonical` — class instances are rejected, not collapsed.** A
+  `Types.ObjectId`, `Buffer` or hydrated Mongoose subdocument has no own
+  enumerable keys, so it digested as `{}` and two manifests referencing DIFFERENT
+  documents hashed identically — the integrity check reported "unchanged".
+  `canonicalJson` now throws `CanonicalizeError` for any object whose prototype is
+  neither `Object.prototype` nor `null`. Plain objects, null-prototype objects,
+  arrays and `Date` are unaffected.
+
+### Changed — the publish smoke gate now sweeps the whole export map
+
+`tests/smoke/smoke.mjs` derives its subpath list from `package.json#exports`
+(44/44) instead of a hand-maintained list that covered 23, and asserts each entry
+declares both `types` and `default` and that both files exist. The gate runs in
+`prepublishOnly`; previously a subpath that lost its tsdown entry would have
+published an export map pointing at a missing file with the gate printing OK.
+
+## [0.19.0] - 2026-08-01
+
+### Added — `./monetization`: canonical pricing classification contract
+
+- **`@classytic/primitives/monetization`** — the single home for how a product
+  is sold and priced, so the classification survives end-to-end across catalog,
+  revenue, order, and entitlement kernels without each maintaining its own enum.
+- **`Monetization`** discriminated union — `FreeMonetization | OneTimeMonetization |
+  SubscriptionMonetization | BundleMonetization | UsageMonetization`, each
+  carrying its own nested pricing shape.
+- **`MONETIZATION_KINDS`** + **`MonetizationKind`** + **`isMonetizationKind`** —
+  exhaustive kind list with runtime guard.
+- Type-guard predicates: **`isFreeMonetization`**, **`isOneTimeMonetization`**,
+  **`isSubscriptionMonetization`**, **`isBundleMonetization`**,
+  **`isUsageMonetization`**.
+- Supporting types: `PriceTier`, `OneTimePricing`, `SubscriptionPlan`,
+  `UsageTier`, `UsageRating`, `DurationUnit`, `MoneyByCurrency`.
+
+### Added — `./proration`: pure proration arithmetic
+
+- **`@classytic/primitives/proration`** — timezone-agnostic, policy-free
+  proration math. Holds no commercial policy (no upgrade/downgrade semantics,
+  no discount or tax logic) — those live in the contract kernel.
+- **`periodProgress(input)`** — fraction of a billing period consumed given
+  start, end, and current instants as `Date`. Returns `PeriodFraction`.
+- **`splitByPeriodFraction(amount, fraction, granularity?)`** — split a `Money`
+  amount into consumed/remaining parts; `granularity: 'whole_day'` (default) or
+  `'exact'`.
+- **`allocateMoneyByFraction(amount, fraction)`** — proportional allocation
+  using the largest-remainder allocator so `consumed + remaining === amount`
+  exactly (no penny drift).
+- **`ProrationError`** + **`ProrationErrorCode`** (`INVALID_PERIOD` |
+  `INVALID_FRACTION`) — typed errors for out-of-range inputs.
+- Supporting types: `PeriodFraction`, `PeriodProgressInput`,
+  `ProrationGranularity`.
+
+## [0.18.0] - 2026-07-29
+
+### Added — `./subject`: `SubjectRef` polymorphic identity pointer
+
+- **`@classytic/primitives/subject`** — `SubjectRef` (`{ subjectModel, subjectRef }`) — WHO a record is about, polymorphically. The pairing spans every kernel that holds data about a person without owning them (attendance, access, loyalty, credentials). Deliberately kept distinct from `ExternalRef` (provenance — "came from") even though the shapes are identical: a grant carries both at once and they mean different things.
+- **`buildSubjectRef(model, id)`** — returns `SubjectRef | null`; `null` when `id` is absent/blank, because a malformed binding silently matches nothing (an entitlement nobody holds, a credential no revoke can find) rather than carrying `"undefined"` as a string.
+- **`SubjectRef`** type exported from `@classytic/primitives/subject`.
+
+### Added — `./state-diagram`: Mermaid diagram from a `StateMachine`
+
+- **`@classytic/primitives/state-diagram`** — `renderStateDiagram(machine, opts?)` — generates a `stateDiagram-v2` Mermaid string from the same `StateMachine` object the runtime CAS enforces. Removes the class of drift where a hand-drawn prose comment diverges from the actual transition table. Pure: string in, string out, no I/O.
+- **`StateDiagramOptions`** — `title?`, `highlight?` (mark specific states), `noteOn?` (per-state annotations).
+
+### Added — `./suspension`: time-bounded suspension/pause policy
+
+- **`@classytic/primitives/suspension`** — pure suspension policy computation (no persistence). Models a pause the subject is ENTITLED to take, bounded by an annual allowance, that ends by itself — distinct from `/hold` (an indefinite blocker needing manual resolution).
+- **`evaluateSuspension(policy, history, opts)`** — always returns `autoResumeAt` (the instant the pause MUST end) so a sweep can resume it whether or not the subject ever comes back. Covers: gym membership freeze, subscription pause, unpaid leave.
+- **`SuspensionPolicy`** — `maxDaysPerPeriod`, `minDurationDays`, `period` ('annual' | 'rolling').
+- **`SuspensionEntry`** — a historical pause record (`startedAt`, `endedAt?`, `reason?`).
+- **`SuspensionResult`** — `autoResumeAt`, `daysUsed`, `daysRemaining`, `status`.
+
+### Added — `OutboxStore.transactionalSave` (`/outbox`)
+
+- **`OutboxStore.transactionalSave?: boolean`** — declares whether `save` enlists `options.session` in the caller's transaction. Absent/`false` = "assume NOT atomic". Set `true` ONLY when `save` passes the session to the same transactional resource the caller writes (e.g. a mongoose model on the same connection). Callers that need atomicity (`@classytic/access`'s `atomically()`) check this flag and fail at boot rather than claiming a guarantee they cannot keep.
+- **`MemoryOutboxStore.transactionalSave = false`** — explicitly declared; an in-memory store cannot participate in a Mongo transaction, so its write survives a rollback.
+
+### Changed — `shelf-life.ts` deduplicates `addDays`
+
+- `addDays` is now imported from `./calendar` instead of being re-defined inline. No behavior change.
+
+## [0.16.0] - 2026-07-27
+
+### Added — `assertApproved` + `CHAIN_INCOMPLETE` (`/approval`)
+
+- **`ApprovalErrorCode` gains `'CHAIN_INCOMPLETE'`** — the single canonical
+  error code for "finalize/commit attempted before the chain reached `approved`".
+  Hosts route it through `rethrowApprovalError` to the `approval.chain_incomplete`
+  wire code so all commit/post/finalize boundaries raise the same structured error.
+- **`assertApproved(chain, message?)`** — canonical finalize-time gate: asserts
+  the chain has reached `approved`, else throws `ApprovalError('CHAIN_INCOMPLETE')`.
+  A `null`/absent chain passes (no chain required). Use at every journal-entry
+  post, PO approve, transfer dispatch, etc. to replace per-domain hand-rolled
+  divergent codes with one definition.
+
+Purely additive — no existing exports changed.
+
+## [0.15.0] - 2026-07-25
+
+### Added — `/retention`: `PurgeEvidence` value object
+
+- **New `@classytic/primitives/retention` subpath** exporting **`PurgeEvidence`**
+  — the persisted evidence record for a GDPR / data-retention purge operation
+  (proof of what was purged, over what scope, by whom, why, and whether
+  analytical measures were retained). Shape: `{ id, subject: { ref, model? },
+  scope, strategy: 'hard'|'soft'|'anonymize', measuresRetained, processed,
+  occurredAt, actor: { ref, kind: 'user'|'system'|'service' }, reason,
+  legalBasis?, status: 'completed'|'partial'|'failed', operationId?,
+  startedAt?, completedAt?, results?: PurgeResourceResult[],
+  verification?: PurgeVerificationSummary }`.
+- **`createPurgeEvidence(input)`** — pure builder (auto-fills `id` + `occurredAt`
+  when absent, defaults `status` to `'completed'`). **`isPurgeEvidence(value)`**
+  type guard + **`assertPurgeEvidence(value)`** throwing assertion.
+
+### Added — `/canonical`: strict deterministic JSON + integrity digest
+
+- **New `@classytic/primitives/canonical` subpath** — `canonicalJson(value)`
+  produces a deterministic string (keys sorted recursively, `Date` → `{"$date":"<iso>"}`,
+  strict rejection of `undefined`/`NaN`/`BigInt`/`Map`/`Set`/cycles with
+  `CanonicalizeError`). **`sha256Hex(input)`** and **`canonicalDigest(value)`**
+  (= `sha256Hex(canonicalJson(value))`) — integrity checksums, not tamper-proof
+  seals. Uses `node:crypto` directly (same precedent as `/otp`).
+
+### Added — `/event-infra`: `propagateHandlerErrors` option
+
+- **`InProcessEventBusOptions.propagateHandlerErrors?: boolean`** (default
+  `false` — existing error-isolating behavior unchanged). When `true`, `publish`
+  runs handlers sequentially and rethrows the first handler error to the
+  publisher — for single-consumer projection/relay lanes where a failed handler
+  must not be silently acked. `publishMany` maps each event's failure into its
+  `PublishManyResult` entry instead of rejecting the batch.
+
+## [0.13.0] - 2026-07-22
+
+### Added — `/unit-cost-rate`: exact per-unit monetary rates (bigint-safe)
+
+- **`@classytic/primitives/unit-cost-rate`** — fractional per-unit monetary
+  rates stored as scaled integers (`scaledAmount = minorPerUnit × RATE_SCALE`,
+  `RATE_SCALE = 1_000_000`). Solves the WAC / FIFO valuation problem where
+  dividing a total by a fractional quantity yields a sub-minor-unit rate
+  (100 paisa ÷ 3 units = 33.333… paisa/unit) that must not lose precision.
+  Exports:
+  - `unitCostRate(minorPerUnit, currency)` — direct rate constructor
+  - `unitCostRateFromTotal(totalMinor, quantity, currency)` — WAC/FIFO
+    constructor; bigint-safe, QTY_SCALE grid for fractional quantities
+  - `extendedAmount(rate, qty, mode?)` — the ONE rate→total rounding boundary;
+    bigint multiplication to keep large amounts exact; default `'half-even'`
+    (banker's rounding), `'half-up'` available for legacy parity
+  - `rateMinorPerUnit(rate)` — display helper (fractional number, NOT for math)
+  - `isUnitCostRate(value)` — structural type guard
+  - `UnitCostRate<C>` interface, `UnitCostRateError`, `UnitCostRateErrorCode`,
+    `RoundingMode`, `RATE_SCALE`
+  - For distributing a top-down charge across lines with no lost minor unit,
+    compose with `allocate` from `@classytic/primitives/split-allocation`.
+
+### Changed — `/outbox` is now the canonical contract owner
+
+- **`@classytic/primitives/outbox` owns the transactional-outbox contract.**
+  Previously this file was documented as a mirror of `@classytic/arc`'s copy
+  ("arc is source of truth, keep bit-identical") while arc's docs said the
+  opposite — a manual-sync drift risk. Settled: primitives owns pure
+  cross-package contracts (events + outbox); arc owns runtime behavior
+  (`EventOutbox` relay, stores, backoff, Fastify integration) and, from
+  arc 2.24+, imports + re-exports this contract instead of duplicating it.
+- **Cross-package `instanceof` now works.** Because arc re-exports these exact
+  `OutboxOwnershipError` / `InvalidOutboxEventError` classes, a store built
+  against primitives throws the same class identity arc's relay catches.
+  Under the old duplication, arc's `instanceof` checks silently missed
+  primitives-thrown ownership errors.
+- No export names or signatures changed — contract surface is identical.
+
+## [0.11.0] - 2026-07-08
+
+### Added — eKYC vocabulary + secure OTP primitive
+
+- **`@classytic/primitives/identity-verification`** — cross-package eKYC data shapes:
+  `IdentityDocumentType`, `VerificationCheck`, `VerificationLevel`, `SubjectKind`,
+  `IdentitySubject`, `DocumentReference`, `IdentityProviderCapabilities`,
+  `VerificationStatus`, `CheckResult`, `CreateVerificationParams`,
+  `VerificationSession`, `VerificationWebhookEvent`. Pure helpers: `isVerified`,
+  `checksSatisfy`, `deriveStatus`. The identity twin of `payment-gateway` — shapes
+  here, contracts in engines, provider implementations anywhere. Serves `@classytic/esign`
+  (IdentityBridge port), `@classytic/payee` (ComplianceBridge), and PSP merchant KYC.
+- **`@classytic/primitives/otp`** — secure OTP challenge lifecycle. Guarantees: code
+  never persisted (only HMAC-SHA256 digest stored), constant-time `timingSafeEqual`
+  comparison, rejection-free `randomInt` CSPRNG (no modulo bias), bounded attempt
+  counter, TTL expiry, single-use verified state. Exports: `createOtpChallenge`,
+  `verifyOtpChallenge`, `isOtpVerified`, `isOtpActive`. Delivery stays in
+  `@classytic/notifications` — this primitive owns only the verify half.
+
+## [0.9.1] - 2026-07-04
+
+### Added
+
+- **`CURRENCY_PATTERN` exported from `/currency`** — the ISO 4217 regex
+  (`/^[A-Z]{3}$/`) behind `isCurrencyCode` / `toCurrencyCode`. Consumers
+  embed the SAME pattern in JSON-Schema-representable validators (zod
+  `.regex`, Mongoose `match`, OpenAPI `pattern`) instead of hand-rolling
+  copies — `.refine(isCurrencyCode)` validates at runtime but is dropped
+  by `z.toJSONSchema`, silently losing the constraint from generated API
+  docs.
+
+## [0.9.0] - 2026-07-02
+
+### Added — `/timezone`: IANA resolution + civil dates (the door for `/calendar`'s escape hatch)
+
+`/calendar` owns wall-clock ARITHMETIC over a fixed `UtcOffsetMinutes` and
+tells hosts to "resolve the offset for the instant and pass it in" — but the
+stack had no shared resolver, so consumers hand-rolled `Intl.DateTimeFormat`
+(three divergent copies found: catalog ×2, be-prod POS). `/timezone` is that
+resolver — zero-dep (Intl IS the ICU tz database), Temporal-shaped names for
+a future mechanical migration:
+
+- **`zoneOffsetMinutes(instant, zone)`** — per-instant, DST-exact offset;
+  composes directly with `/calendar`: `startOfMonth(i, zoneOffsetMinutes(i, z))`.
+- **`localTimeParts(instant, zone)`** — `{ isoWeekday, hour, minute,
+  minutesOfDay }` in the zone; what pricing windows / slot grids /
+  working-hours checks evaluate against (canonicalizes catalog's resolvers).
+- **`CivilDate`** (branded `'YYYY-MM-DD'`) + `civilDate` / `isCivilDate` /
+  `civilDateOf(instant, zone)` / `civilDateToInstant(cd, zone, time?)`
+  (two-pass DST-safe) / `addCivilDays` / `civilDaysBetween` — the type for
+  hotel nights, POS business dates, VAT filing years. Lexicographic ==
+  chronological, so it's a range-queryable Mongo string key.
+- **`isValidTimeZone(zone)`** — boot-time host-config validation;
+  **`listTimeZones()`** + **`zoneOffsetLabel(instant, zone)`** — the data
+  source + label for Google-Calendar-style zone pickers, straight from ICU
+  (`Intl.supportedValuesOf`), no bundled list to go stale.
+- `TimeZoneError`; formatter instances cached per zone.
+
+No arithmetic here (that stays in `/calendar`) — resolution only.
+
+## [0.7.2] - 2026-05-30
+
+### Added — `BankTransaction` per-account routing + pending lifecycle fields
+
+Three additive optional fields on `BankTransaction` (`/bank-transaction`):
+
+- **`sourceAccountId?: string`** — the vendor-side account this row belongs to
+  (Plaid `account_id`, Xero `BankAccount.AccountID`). Sync feeds return
+  transactions spanning multiple accounts in one batch; this lets consumers
+  route each row to the right account per-transaction (multi-account sync)
+  instead of forcing the whole batch onto one account. Statement formats
+  (single-account files) leave it undefined — the account lives on
+  `BankStatement.account`.
+- **`pending?: boolean`** — the entry hasn't posted/cleared yet (real-time
+  feeds like Plaid `pending: true`; statement formats leave it undefined).
+  Consumers must treat pending rows as provisional — don't reconcile/post
+  them as final.
+- **`supersedesExternalId?: string`** — set on a POSTED row that replaces an
+  earlier PENDING one, carrying the pending row's `externalId` (Plaid
+  `pending_transaction_id`). Pending and posted rows have different
+  `externalId`s, so consumers MUST use this back-pointer to drop/replace the
+  superseded pending row on ingest — otherwise both persist as duplicates.
+
+Backward-compatible (both optional). Closes a fidelity gap where Plaid's
+pending state was dropped at the canonical boundary, so downstream consumers
+(`@classytic/fin-io` Plaid mapper, `@classytic/revenue` import) couldn't
+distinguish provisional rows or dedupe pending→posted.
+
+## [0.7.1] - 2026-05-26
+
+### Added — workflow + scheduling + composition primitives
+
+Four new flat subpaths (additive):
+
+- **`/status-history`** — append-only status transition log primitive.
+- **`/condition`** — declarative condition expressions for workflow gates.
+- **`/mixin`** — composition helpers for primitive object assembly.
+- **`/sla-policy`** — SLA policy definitions paired with the existing `/sla` runtime.
+
+### Added — payment event coverage
+
+Expanded `PAYMENT_EVENT_TYPE` catalogue to cover auth/capture and dispute
+lifecycles: `AUTHORIZED`, `CAPTURED`, `AUTH_VOIDED`, `DISPUTED`,
+`DISPUTE_WON`, `DISPUTE_LOST`, `SETTLED`.
+
+`PaymentRefundedPayload` and `PaymentReversedPayload` now carry
+`originalAmount` (and `reversedAmount` / `isPartial` on reversed) so
+consumers can detect partial vs full operations without a follow-up
+Payment lookup.
+
+## [0.7.0] - 2026-05-26
+
+### Added — payment domain primitives
+
+Three new subpaths shared across every payment-aware package
+(`@classytic/invoice`, `@classytic/revenue`, `@classytic/order`, future
+POS / AP / subscription packages). Each ships under a flat subpath so
+consumers tree-shake what they don't import.
+
+- **`/payment-method-kind`** — `PaymentMethodKind` universal vocabulary
+  (kind vs host-registered `methodCode`) so packages can speak `kind`
+  while hosts speak `code`.
+- **`/payment-allocation-status`** — orthogonal allocation +
+  reconciliation status enums for payment legs (`unallocated`,
+  `partial`, `allocated`, `over_allocated`, plus reconciliation states).
+- **`/payment-events`** — pure TS payload contracts for payment domain
+  events. Wire-only (no Zod, no runtime). Anchored on `Money` +
+  `PaymentMethodKind`; causation/correlation stay on `EventMeta`.
+
+No breaking changes.
+
+## [0.6.0] - 2026-05-17
+
+### Added — five new primitives for CRM / workflow domains
+
+Five new primitives covering gaps the existing surface didn't address.
+Each ships under its own flat subpath import (no breaking changes).
+
+- **`/phone`** — `PhoneNumber` value object + `parsePhone(input)` that
+  normalizes free-form input to E.164, captures the country calling code
+  as a discrete field, and reports failures via `Result<PhoneNumber,
+  PhoneError>`. Zero external deps; ITU-T E.164 prefixes inlined. Hosts
+  that want region-of-issue / line-type detection layer on top with
+  libphonenumber-js.
+
+- **`/status-history`** — Generic, immutable status-change log:
+  `StatusChangeEntry<TStatus>`, `appendStatus()`, `timeInStatus()`,
+  `lastTransitionTo()`. Captures `durationInPriorMs` at write time so
+  funnel-velocity dashboards ("avg time in Qualification") work without
+  re-walking timestamps. Pairs naturally with `/state-machine`.
+
+- **`/condition`** — Tiny JSON-serializable predicate DSL: `FieldCondition`
+  (`{ field, op, value }`) composed via `AllCondition` / `AnyCondition`
+  / `NotCondition`, plus a total `evaluate(condition, target)` and a
+  `validateCondition()` for boot-time schema checks. Powers SLA scope
+  filters, drip "if opened/clicked" branches, automation triggers, and
+  permission row-scopes — anywhere a predicate needs to round-trip
+  through a database.
+
+- **`/mixin`** — Additive type composition without subclassing:
+  `withMixin<T, K, M>()` / `getMixin()` / `hasMixin()` / `withoutMixin()`.
+  Lets a `Contact` carry per-domain payloads (`mixins.customer`,
+  `mixins.lead-fit`) on a typed `mixins` namespace without bloating the
+  base type. Plain object composition — no decorators, no global
+  hierarchy, no proxies.
+
+- **`/sla-policy`** — Higher-order SLA layered on the existing `/sla`:
+  priority matrix (`urgent` / `high` / `normal` → durations), first-
+  response vs rolling-response semantics, and working-hours window
+  (weekdays + start/end minute + holidays). `defineSLAPolicy()` validates
+  at boot; `evaluateSLAStatus()` produces `{ kind, responseBy,
+  remainingMs, breached }` from a policy + per-entity inputs. The
+  existing simple `/sla` interface is unchanged — `sla-policy` derives
+  concrete `SLA` instances from a policy + priority and delegates to it.
+
+### Changed — source-tree reorganization (no public-surface impact)
+
+The 24 primitive source files moved from a flat `src/*.ts` layout into
+six category folders so the package is self-documenting at the file
+system level:
+
+```
+src/
+  money/           money, currency, split-allocation, bank-transaction, payment-gateway
+  identity/        person, address, phone
+  scheduling/      period, cadence, sla, sla-policy
+  workflow/        state-machine, status-history, approval, hold, condition
+  events/          events, outbox
+  composition/     mixin, reference, context, brand, result
+```
+
+**Public subpath imports are unchanged** — `@classytic/primitives/money`,
+`@classytic/primitives/sla`, etc. still resolve. tsdown emits each entry
+to `dist/<name>.mjs` regardless of source nesting. Existing consumers
+upgrade with `npm install @classytic/primitives@^0.6.0` and a no-op
+diff.
+
+### Dependencies
+
+Still zero runtime deps. Node 22+.
+
+## [0.5.0] - 2026-05-05
+
+### Added — `/bank-transaction` subpath
+
+Canonical bank-feed shapes shared across `@classytic/fin-io` (parsers
+produce these), `@classytic/revenue` (`transactionRepository.import()`
+consumes them), and `@classytic/ledger` (`bankStatementMapper`
+consumes them). Single source of truth for the cross-package
+vocabulary; eliminates the structural mirrors fin-io and revenue
+were carrying.
+
+Exports: `BankTransaction`, `BankCounterparty`, `BankAccount`,
+`BankStatement`, `BankStatementSource`, `BankImportReport`,
+`BankImportRowError`. All anchored on `Money` from `/money` (integer
+minor units, `number`).
+
+### Added — `/payment-gateway` subpath
+
+Canonical payment-gateway data shapes. **Lets any provider package
+(Stripe, Razorpay, SSLCommerz, PayPal, bKash, Nagad, manual, …)
+implement against primitives' types without depending on revenue's
+heavyweight runtime.** Provider authors peer-dep on
+`@classytic/primitives` only.
+
+Exports: `CreateIntentParams`, `PaymentIntent`, `PaymentResult`,
+`RefundResult`, `WebhookEvent`, `ProviderCapabilities`. All `amount`
+fields are `Money` from `/money` (no flat `amount: number, currency:
+string` pairs — currency travels inside Money).
+
+The `PaymentProvider` abstract class — the *contract* every provider
+must satisfy — stays in `@classytic/revenue`. Revenue is the data
+engine that consumes providers; provider packages live elsewhere.
+
+### Architectural intent
+
+Revenue 3.0 + primitives 0.5 + fin-io 0.3 establish a clean
+three-tier seam:
+
+  1. **primitives** — pure data shapes (no runtime, no deps). Owns
+     `Money`, `BankTransaction`, `PaymentIntent`, every cross-
+     package vocabulary type.
+  2. **revenue** — the engine. Owns the abstract contracts
+     (`PaymentProvider`, `BankFeedProvider`), state machines, repos.
+     Consumes primitives shapes.
+  3. **provider packages** (`revenue-stripe`, `revenue-razorpay`,
+     `@classytic/fin-io/adapters/revenue`, …) — implementations.
+     Peer-dep on primitives. Optional peer-dep on revenue for the
+     abstract base class.
+
+This eliminates the cycle where every provider package transitively
+pulled mongoose + mongokit just to know what a `PaymentIntent`
+looked like. Tree-shaking, build times, and dep-graph honesty all
+improve.
+
+### Migration
+
+Consumers MUST import from primitives subpaths. No re-exports from
+revenue or fin-io (PACKAGE_RULES P2):
+
+```ts
+// Before
+import type { PaymentIntent, CreateIntentParams } from '@classytic/revenue';
+import type { CanonicalTransaction, Money } from '@classytic/fin-io';
+
+// After
+import type {
+  PaymentIntent,
+  CreateIntentParams,
+} from '@classytic/primitives/payment-gateway';
+import type { BankTransaction } from '@classytic/primitives/bank-transaction';
+import type { Money } from '@classytic/primitives/money';
+```
+
+Provider helper classes (`PaymentIntent`, `PaymentResult`,
+`RefundResult`, `WebhookEvent`) are deleted from revenue — they were
+trivial wrappers around plain objects. Provider implementations now
+return plain object literals matching the interface.
+
+`CreateIntentParams.amount` is now `Money` (`{ amount: number,
+currency: string }`), not the flat `amount + currency` pair. Revenue's
+repository wraps internally:
+
+```ts
+const intent = await provider.createIntent({
+  amount: { amount: params.amount, currency },
+  metadata: params.metadata,
+  ...params.paymentData,
+});
+```
+
+## [0.3.1] - 2026-05-02
+
+### Added — state-machine ↔ mongokit/sqlitekit integration
+
+State-machine primitives now compose directly with the kit-level CAS surface shipped in mongokit 3.13 and the sqlitekit follow-up. Three additions:
+
+- **`StateMachine.validTargets(from)`** — every status the aggregate can move TO from `from`. Useful for UI dropdowns ("which actions are legal right now?") and for fanning a single source state into a multi-source CAS via a repo's `claim({ from: machine.validTargets(current), to })`.
+- **`StateMachine.validSources(to)`** — every status that can transition INTO `to` (reverse adjacency). Pair with a repo's multi-source `claim({ from: machine.validSources(to), to })` when the target is fixed but the caller doesn't yet know the current state — the canonical "cancel from any non-terminal" / "error from any in-flight" pattern. Pre-computed at definition time, O(1) lookup, frozen arrays safe to expose by reference.
+- **`assertAndClaim(machine, repo, id, args)`** — pairs `machine.assertTransition` (sync domain check) with `repo.claim()` (atomic CAS) in one call. The canonical state-machine-backed CAS pattern. Skipping either layer leaves a hole: skip `assertTransition` and bad transitions reach storage; skip `claim` and concurrent writers race.
+
+```ts
+import { defineStateMachine, assertAndClaim } from '@classytic/primitives/state-machine';
+
+const ORDER_MACHINE = defineStateMachine<OrderStatus>({
+  name: 'Order',
+  transitions: {
+    draft:     ['approved', 'cancelled'],
+    approved:  ['shipped', 'cancelled'],
+    shipped:   [],
+    cancelled: [],
+  },
+});
+
+// Single-source:
+const updated = await assertAndClaim(ORDER_MACHINE, orderRepo, orderId, {
+  from: 'draft',
+  to: 'approved',
+  patch: { approvedAt: new Date() },
+  options: { organizationId: ctx.organizationId, session },
+});
+
+// Multi-source via reverse adjacency — "cancel from any legal predecessor":
+await assertAndClaim(ORDER_MACHINE, orderRepo, orderId, {
+  from: ORDER_MACHINE.validSources('cancelled'), // ['draft', 'approved']
+  to: 'cancelled',
+  patch: { cancelledAt: new Date() },
+  options: { organizationId: ctx.organizationId },
+});
+```
+
+#### `ClaimableRepo<TDoc>` — structural dependency on the kit
+
+`assertAndClaim` accepts any object matching `ClaimableRepo<TDoc>` — a structural type that mirrors the mongokit 3.13 `Repository.claim()` signature. Primitives doesn't take a hard peer dep on mongokit; consumers wire the kit they want. Sqlitekit's `claim()` matches the same shape (verified by mongokit/sqlitekit cross-conformance).
+
+```ts
+export interface ClaimableRepo<TDoc> {
+  claim(
+    id: string,
+    transition: {
+      field?: string;
+      from: unknown | readonly unknown[];
+      to: unknown;
+      where?: Record<string, unknown>;
+    },
+    patch?: Record<string, unknown>,
+    options?: Record<string, unknown>,
+  ): Promise<TDoc | null>;
+}
+```
+
+#### Why a 0.3.1 (not 0.4.0)
+
+Purely additive — no removals, no signature changes. `defineStateMachine` callers from 0.3.0 see two new methods on the returned `StateMachine` (`validTargets`, `validSources`) and one new free function (`assertAndClaim`). Backwards-compatible.
+
+## [0.3.0] - 2026-04-29
+
+### Removed — `/errors` subpath (BREAKING)
+
+- Deleted `@classytic/primitives/errors` (`ErrorContract`, `ErrorDetail`, `ErrorCode`, `ERROR_CODES`).
+- Errors are infrastructure-shaped (HTTP-coupled, repository-coupled, RFC 7807-shaped). Moved to `@classytic/repo-core/errors` where they sit next to `pagination`, `tenant`, `cache`, `hooks`, `schema` — every other repository contract — and join the existing `HttpError` throwable contract there. Single canonical home for both the wire (`ErrorContract`) and throwable (`HttpError`) shapes.
+
+#### Migration
+
+```diff
+- import type { ErrorContract, ErrorDetail, ErrorCode } from '@classytic/primitives/errors';
+- import { ERROR_CODES } from '@classytic/primitives/errors';
++ import type { ErrorContract, ErrorDetail, ErrorCode } from '@classytic/repo-core/errors';
++ import { ERROR_CODES } from '@classytic/repo-core/errors';
+```
+
+Field shapes and code values are unchanged — pure relocation.
+
+### Removed — `/tenant` subpath (BREAKING)
+
+- Deleted `@classytic/primitives/tenant` (`TenantConfig`, `TenantStrategy`, `TenantFieldType`, `DEFAULT_TENANT_CONFIG`, `resolveTenantConfig`, `ResolvedTenantConfig`).
+- Tenant scope is **infrastructure-shaped** (describes how queries get scoped to a database / repository), not a domain primitive like Money or Address. Moved to `@classytic/repo-core/tenant` where it sits next to `context`, `filter`, `hooks`, `schema`, `cache` — every other repository contract.
+- Mongokit and sqlitekit already peer-dep `@classytic/repo-core`, so consuming `TenantConfig` from there costs zero extra peer deps for those kits.
+
+#### Migration
+
+```diff
+- import type { TenantConfig } from '@classytic/primitives/tenant';
+- import { resolveTenantConfig, DEFAULT_TENANT_CONFIG } from '@classytic/primitives/tenant';
++ import type { TenantConfig } from '@classytic/repo-core/tenant';
++ import { resolveTenantConfig, DEFAULT_TENANT_CONFIG } from '@classytic/repo-core/tenant';
+```
+
+Field shapes, runtime semantics, and defaults are unchanged. The custom-strategy escape hatch (`strategy: 'custom'` + `resolve: (ctx) => filterShape`) works identically.
+
+### Removed — `/pagination` subpath (BREAKING)
+
+- Deleted `@classytic/primitives/pagination` (`PageParams`, `KeysetParams`, `SortSpec`, `SortDirection`, `OffsetPage<T>`, `KeysetPage<T>`, `AggregatePage<T>`, `emptyOffsetPage`, `emptyKeysetPage`).
+- Pagination shapes are owned by **`@classytic/repo-core/pagination`** — that package ships both the types AND the cursor encoding/skip-math/keyset-validation algorithms. Primitives' duplicate type-only declarations existed before the contract was canonised in repo-core; keeping them invited drift.
+
+### Migration
+
+```diff
+- import type { OffsetPage, KeysetPage, AggregatePage, SortSpec } from '@classytic/primitives/pagination';
++ import type {
++   OffsetPaginationResult,
++   KeysetPaginationResult,
++   AggregatePaginationResult,
++   SortSpec,
++ } from '@classytic/repo-core/pagination';
+```
+
+Note: shape names also changed (`OffsetPage` → `OffsetPaginationResult`, etc.) to match repo-core's naming. Field names align with what mongokit, sqlitekit, and arc already produced — those packages always returned repo-core shapes; primitives' types were a parallel, never-quite-aligned vocabulary.
+
+If you also need HTTP wire envelopes (`{ success: true } & Result`), repo-core 0.3.0 ships `OffsetPaginationResponse`, `KeysetPaginationResponse`, `AggregatePaginationResponse`, `BareListResponse`, `PaginatedResponse`, plus a `toCanonicalList()` runtime normalizer.
+
+### Why now (and why this is the only breaking change in this release)
+
+Two layers of pagination shapes drifted across the org: primitives declared one set, repo-core declared a different set, mongokit + sqlitekit + arc all consumed repo-core's. Removing primitives' set is purely subtractive — nobody internally imports from `@classytic/primitives/pagination`, and external consumers get a clear migration path to repo-core. No new ground to break later.
+
+## [0.2.0]
+
+### Added
+
+- **`/state-machine` subpath** — `defineStateMachine<TStatus>({ name, transitions, errorFactory? })`
+  declarative state-transition primitive. Replaces hand-rolled
+  `if (status === 'X') throw` blocks across kernels:
+
+  ```typescript
+  import { defineStateMachine } from '@classytic/primitives/state-machine';
+
+  const ORDER_MACHINE = defineStateMachine<OrderStatus>({
+    name: 'Order',
+    transitions: {
+      draft: ['approved', 'cancelled'],
+      approved: ['shipped', 'cancelled'],
+      shipped: [],
+      cancelled: [],
+    },
+  });
+
+  ORDER_MACHINE.assertTransition('order-1', 'draft', 'approved'); // OK
+  ORDER_MACHINE.assertTransition('order-1', 'shipped', 'draft');  // throws
+  ORDER_MACHINE.canTransition('draft', 'shipped');                // false
+  ORDER_MACHINE.isTerminal('shipped');                            // true
+  ```
+
+  Generic over the status string-union — `Record<TStatus, readonly TStatus[]>`
+  on `transitions` enforces exhaustiveness at the call site. Kernels with their
+  own typed transition error class (e.g. flow's `InvalidTransitionError`) wire
+  it via `errorFactory` so the thrown type stays domain-consistent. Standalone
+  callers get the bundled `IllegalTransitionError` (carries `entityType` /
+  `entityId` / `from` / `to` + `code: 'illegal_transition'` + `status: 422`).
+
+  First consumer: `@classytic/flow` `PROCUREMENT_MACHINE` (procurement-order
+  approve / cancel transitions).
+
+## [0.1.1]
+
+### Added
+
+- **`MinorUnits<C>` branded type on `/money`** — wire-format monetary amount
+  (integer minor units, branded with phantom currency) for HTTP/JSON fields
+  where a full `Money` struct is overkill. Erases at runtime; `MinorUnits<'BDT'>`
+  vs `MinorUnits<'USD'>` collide at compile time. Pair with structured `Money`
+  for in-process arithmetic. Constructor: `MinorUnits(amount, currency?)`.
+
+### Changed
+
+- **`/tenant` `contextKey` cascade default** — `resolveTenantConfig({ tenantField: 'branchId' })`
+  now defaults `contextKey` to `'branchId'` instead of `'organizationId'`.
+  When a host renames `tenantField`, their `OperationContext` carries the id
+  under the same key in the overwhelming majority of cases; mirroring the
+  rename is the least-surprise default. Explicit `contextKey` still overrides
+  (set when the doc field and context key genuinely diverge). Existing call
+  sites that supplied both `tenantField` and `contextKey` are unaffected.
+
+## [0.1.0]
+
+Initial release. Framework-agnostic TypeScript primitives shared across
+Classytic packages. Zero runtime dependencies. Node 22+. ESM only.
+260 tests (unit + integration) across 20 files.
+
+All subpaths are importable independently; there is **no root barrel
+re-export** — consumers must use subpath imports (`@classytic/primitives/money`,
+`/events`, `/tenant`, …) to keep the Node ESM loader + TypeScript compiler
+from walking the full module graph on any convenience import.
+
+### Subpaths
+
+- **`/money`** — `Money` type + integer-minor-unit arithmetic (`money`,
+  `addMoney`, `subtractMoney`, `multiplyMoney`, `sumMoney`, `negateMoney`,
+  `absMoney`, `compareMoney`, `equalsMoney`, `isZeroMoney`, `isPositiveMoney`,
+  `isNegativeMoney`, `isMoney`, `fromMajor`, `toMajor`); `CurrencyMismatchError`.
+- **`/currency`** — `CurrencyCode` brand; `CURRENCIES` table; `MINOR_UNIT_FACTOR`,
+  `minorUnitFactor`, `isCurrencyCode`, `toCurrencyCode`.
+- **`/address`** — `Address`, `ContactAddress`, `GeoPoint`, `GeoJsonPoint`;
+  `toGeoJsonPoint`, `fromGeoJsonPoint`.
+- **`/period`** — `DateRange`, `Period`; `isDateRange`, `isWithin`,
+  `rangeDurationMs`.
+- **`/pagination`** — `PageParams`, `OffsetPage<T>`, `KeysetPage<T>`,
+  `AggregatePage`, `KeysetParams`, `SortSpec`, `SortDirection`;
+  `emptyOffsetPage`, `emptyKeysetPage`.
+- **`/reference`** — `ExternalRef`, `DocumentRef`, `ObjectIdLike`, `IdLike`;
+  `idToString`, `isExternalRef`.
+- **`/context`** — `OperationContext` (the identity + tracing bag every
+  Classytic package accepts); `ActorRef`.
+- **`/events`** — `EventMeta` (arc v2.9-aligned — 13 fields including
+  `schemaVersion`, `causationId`, `partitionKey`, `source`, `idempotencyKey`,
+  `aggregate`), `DomainEvent<T>`, `EventTransport`, `EventHandler`,
+  `EventLogger`, `DeadLetteredEvent<T>`, `PublishManyResult`; `createEvent`,
+  `createChildEvent` (auto-chains `causationId`, inherits `correlationId` /
+  `userId` / `organizationId` / `source` / `idempotencyKey` — NOT
+  `aggregate`), `matchEventPattern`.
+- **`/outbox`** — transactional outbox **contract** mirrored from arc v2.9:
+  `OutboxStore`, `OutboxWriteOptions`, `OutboxClaimOptions`,
+  `OutboxAcknowledgeOptions`, `OutboxFailOptions`, `OutboxErrorInfo`,
+  `OutboxFailureContext`, `OutboxFailureDecision`, `OutboxFailurePolicy`;
+  `OutboxOwnershipError`, `InvalidOutboxEventError`. **Contract only** —
+  runtime (`EventOutbox`, `MemoryOutboxStore`, `MongoOutboxStore`,
+  `exponentialBackoff`) ships in `@classytic/arc/events` and
+  `@classytic/arc/events/mongo`. Domain packages implement `OutboxStore`
+  without peer-depping arc.
+- **`/tenant`** — `TenantConfig` (with `strategy: 'field' | 'none' | 'custom'`
+  and optional `resolve(ctx)` for custom strategy), `TenantFieldType`,
+  `TenantStrategy`, `ResolvedTenantConfig`; `DEFAULT_TENANT_CONFIG`,
+  `resolveTenantConfig`. Field names match `@classytic/mongokit`'s
+  `MultiTenantOptions` so a resolved config threads into
+  `multiTenantPlugin({ ... })` without translation.
+- **`/result`** — `Result<T, E>`; `ok`, `err`, `isOk`, `isErr`, `unwrap`,
+  `mapResult`, `mapError`.
+- **`/brand`** — `Brand<T, B>`, `Prettify`, `DeepPartial`, `DeepReadonly`,
+  `RequireKeys`, `OptionalKeys`, `NonEmptyArray`.
+- **`/errors`** — `ErrorContract`, `ErrorDetail`, `ErrorCode`, `ERROR_CODES`.
+- **`/split-allocation`** — `allocate(total, subjects, method)` returns a
+  `SplitResult` whose parts sum exactly to `total` via the largest-remainder
+  method. Methods: `equal`, `by-qty`, `by-weight`, `by-volume`, `by-value`,
+  `by-percent`. Deterministic — identical inputs produce identical outputs;
+  ties broken by input order. Handles negative totals by sign-flipping.
+  Exports: `allocate`, `isBalanced`, `SplitAllocationError`, types
+  `SplitMethod`, `SplitSubject`, `SplitPart`, `SplitResult`,
+  `SplitAllocationErrorCode`.
+- **`/approval`** — `ApprovalChain` value object with pure FSM helpers.
+  Supports sequential and parallel chains, per-step quorum
+  (`requiredApprovals`), threshold-gated skips. Any step-level rejection
+  rejects the chain; all steps `approved`|`skipped` approves it. Exports:
+  `createChain`, `applyDecision`, `skipStep`, `nextPendingStep`,
+  `pendingSteps`, `isApproved`, `isRejected`, `isPending`, `decisionCount`,
+  `ApprovalError`, types `ApprovalChain`, `ApprovalStep`, `ApprovalDecision`,
+  `Approver`, `ApprovalStatus`, `ChainOrder`, `ApprovalThreshold`,
+  `CreateChainInput`, `CreateStepInput`, `DecisionInput`, `ApprovalErrorCode`.
+- **`/cadence`** — recurrence spec with pure UTC date arithmetic. Kinds:
+  `daily`, `weekly`, `monthly`, `yearly`, `cron` (opaque — host parses).
+  `dayOfMonth: 31` snaps to the last day of shorter months; Feb 29 snaps
+  to Feb 28 in non-leap years. Exports: `nextOccurrence`,
+  `occurrencesBetween`, `validateCadence`, `CadenceError`, types `Cadence`,
+  `DailyCadence`, `WeeklyCadence`, `MonthlyCadence`, `YearlyCadence`,
+  `CronCadence`, `CadenceKind`, `IsoWeekday`, `CadenceErrorCode`.
+- **`/hold`** — `HoldReason` + collection helpers. Host embeds
+  `HoldReason[]` on documents. Exports: `addHold`, `resolveHold`,
+  `activeHolds`, `resolvedHolds`, `isOnHold`, `hasActiveHoldOfCode`,
+  `HoldError`, types `HoldReason`, `HoldActor`, `HoldActorKind`,
+  `AddHoldInput`, `ResolveHoldInput`, `HoldErrorCode`.
+- **`/sla`** — target duration + breach-detection arithmetic. Policy is
+  carried but not executed — host reads `breachPolicy` (`warn` | `escalate`
+  | `block`) and acts. Exports: `breachedAt`, `remainingMs`, `elapsedMs`,
+  `isBreached`, `consumedFraction`, `validateSLA`, `SLAError`, types
+  `SLA`, `BreachPolicy`, `SLAErrorCode`.
+- **`/person`** — person/contact shapes shared across order, crm, hr.
+  `PersonName`, `ContactInfo`, `EmergencyContact`, `Gender`; formatters
+  `formatFullName`, `formatDisplayName`.
+
+### Design principles
+
+- Pure functions on plain data. No I/O, no Mongo, no events.
+- Immutable return values. Helpers return new collections; originals are never mutated.
+- Deterministic. Same inputs → same outputs. Clock sources injected
+  explicitly so tests don't flake.
+- Errors are typed. Every module ships an `<Name>Error` class with a
+  discriminated `code: <Name>ErrorCode` string union.
+- Subpath-imports only — no root barrel.
+
+### Ownership direction
+
+Arc is the source of truth for event types. Primitives **mirrors** arc
+(`@classytic/arc/events/EventTransport.ts`, `@classytic/arc/events/outbox.ts`)
+verbatim for `EventMeta`, `DomainEvent`, `EventTransport`,
+`DeadLetteredEvent`, `EventLogger`, `PublishManyResult`, and the full
+`OutboxStore` contract surface. Helpers (`createEvent`, `createChildEvent`)
+are allowed to be more defensive but match arc's signatures.
+
+Packages that peer-dep `@classytic/primitives` can write
+`class FooOutboxRepository implements OutboxStore` without pulling arc into
+their dependency graph; hosts import `EventOutbox` + `MongoOutboxStore` from
+arc and wire them against package-provided stores.