DateKeys/CHANGELOG.md

# Changelog

All notable changes to this module are documented here. The project follows
semantic versioning; `v0.x` versions make no API stability promise.

## Unreleased — specification v0.12 draft

Implements the draft v0.12 of the DateKeys Protocol Specification, on the
branch `v0.12` and not approved yet. It fixes what the review of the
implementation of v0.11 found on 2 October 2026, and changes no format: a
reader of v0.11 opens these capsules, and this one opens those of v0.11.
`SpecVersion` stays 0.11 until the author approves the draft.

- **Review of 2 October.** Six adversarial reviews of the implementation of
  v0.11 found nothing blocking: what is signed, the strict Ed25519, the CMS
  and RFC 3161 checks and the envelope hold, and the failures were at the
  edges. What the text of v0.11 already asked for is fixed in the code, as
  this entry says; what it left open or got wrong goes to the draft.
  - A typed nil in `EncryptOptions.AuthorKey`, `CMSSigner` or `Sealer` is an
    error, never a capsule without the signature or the seal that was asked
    for.
  - A panic while evaluating the signature or the seal fails only that part,
    F1 or S2, not both.
  - `OpenOptions.Accept` sees the verdicts after step 17 and before step 18,
    and can refuse to publish the files.
  - `extension.CheckWrite`, the rule of encoders of spec §72: the writers of
    capsules and `.dkk` files refuse `datekeys.note` and `datekeys.capsule`
    outside the arrays where they are registered, or with invalid data.
  - `authorkey.Key.String` and `GoString` hide the secret key, which only
    `Key.Secret` returns. `ParsePublic` refuses a key that is not a point of
    the curve (`ed25519strict.OnCurve`).
  - `Header.UnusableNote` tells a public note that breaks the rules of text
    from no note.
- **Specification.** `spec/DateKeys_Protocol_Specification_v0.12.md`, a draft
  whose §76 lists each change with its case: the names of certificates in the
  verdicts, between « and », shown only with at most 64 code points and no two
  spaces in a row; the authority of each seal in the lines of F6, with the
  warning that nobody checks who issued it; the holder by `givenName` and
  `surname` before the `commonName`, which in the certificates of the FNMT
  carries the NIF, when both have text that is not empty, and the issuer by
  its `commonName` or, without one that has text, its `organizationName`;
  the profile of the certificate field by field; object
  identifiers by their bytes, repeated elements of a SET OF and the edge cases
  of the token; the addresses and the padding of the locator; and the errata
  of §44.1, §55.2, §64, §67 and §76. The CDDL fixes the sizes of the locator.
  `Verdicts.Lines` follows it, and writes the result of a foreign signer in
  Spanish.
- **CMS reader with a profile of its own.** `internal/cms` reads certificates
  field by field, as the profile of §29.10 says, instead of with
  `encoding/asn1` and `crypto/x509`, so that a second implementation reads
  them the same; a certificate that breaks it decides nothing unless a
  `SignerInfo` names it. The text of a name comes only from UTF8String,
  PrintableString, IA5String, TeletexString in ASCII and BMPString without
  surrogates, with nothing removed. The key is RSA with NULL parameters and
  an odd modulus, or EC uncompressed on P-256, P-384 and P-521. Object
  identifiers are compared by the bytes of their DER, so an arc of 2^31 or
  more no longer makes an attribute that decides nothing fail the signature;
  a SET OF may repeat an element, and two copies of a certificate are one; a
  key of another scheme than its algorithm is F2; a `messageImprint` of
  another length is S3; the `crls` of a token decide nothing. `internal/der`
  checks UTCTime and GeneralizedTime in their forms of X.690, a date that
  exists, and millis and micros as minimal INTEGERs.
- **Addresses of the locator.** A reader rejects an address that breaks §44.1
  and uses the others (`Locator.Usable`); a writer never writes one. Addresses
  refuse the special-purpose blocks of IANA, IPv6 outside 2000::/3,
  `localhost` and local names, characters outside RFC 3986, dot segments, and
  a CID that does not decode to version 1 and a multihash. A port with a
  leading zero is refused too, as the draft asks, which also writes the rules
  that the reference applied without a text: segments of 1 to 63 characters
  that neither start nor end with a hyphen, the scheme in lower case, `0x`
  and local names in either case, an IPv4 without leading zeros and a CID in
  canonical form. `ParseInfo` checks
  that the locator is an age file with one tlock stanza for the round of its
  DateKey, `Info.Extension` reads what it writes, and `Info.OpenLocator`
  refuses a nil registry. The errors of `Open` and `Unmarshal` carry no
  normative code.
- **CLI.** `encrypt -sign` shows the author key and the code of
  `AUTHOR_MESSAGE` before it signs (spec §62.1 rule 20).
  `decrypt -expect-author` compares the key of an F4, and writes nothing
  unless it matches. `decrypt` and `inspect` say when a public note is not
  shown because it breaks the rules of text; `inspect -json` gives
  `public_note_unusable`. The lines of the verdicts break at the last space
  that fits, each row after the first behind `  ↳ `, so that the terminal
  never breaks them. `encrypt` reports L as the length of the payload, not of
  the content.
- **Tables for Dart.** `internal/pathrule/gen -dart` writes the Unicode and
  best-fit tables as a Dart library for `datekeys-dart`, the third
  implementation, as `-ts` does for `datekeys-ts`: the same lists and the same
  `TablesDigest`. The tables and the TypeScript module do not change.
- **Test data.**
  - `genfixtures -force` writes the fixtures of v0.10 again with their area
    of 512 bytes, through `EncryptOptions.TestAreaLen`, which needs
    `TestVectors`; it used to rewrite them with 32 KiB and fail halfway.
  - `format3_seal_unsupported` uses `seal_type` 4294967295, reserved for
    tests, instead of `seal_type` 1, which a later version may define;
    `capsule.AlgTest` and `SealTypeTest` name the two values.
  - `security.json` carries the context of a capsule, its commitments and the
    time of its round, and each case the verdicts and the lines of a reader of
    this version: a signature of `alg` 1 that does not verify is F2, a token
    of `seal_type` 2 that is not DER is S2, and new cases give a valid
    signature of `alg` 1 (F4), and `alg` and `seal_type` 4294967295 (F1, S1).
    24 cases.
  - `note.json`, new: 16 public notes, those that pass and those that a
    writer refuses and a reader does not show, each with the text of the rule
    it breaks.
  - New fixtures `format3_unsigned`, the capsule of `format3_signed` without
    its signature, with the same P, and `format3_note`, with a public note:
    26 capsules, 14 of format 3. The records of `format3_signed_cms` and
    `format3_sealed` give the lines of v0.12.
  - `mutations.json` gains 8 mutations of the list of v0.11 of §64: the
    signature of `alg` 1 altered (F2), removed (F0), made again with another
    key (F4 of that key) and transplanted to another capsule (F2); a key of 31
    bytes and a signature of 65 (F1); the area widened to 64 KiB after signing
    (F4, the same `AUTHOR_MESSAGE`); and the public note changed
    (`ERR_HEADER_BINDING`, step 15). The signature of `alg` 1 that does not
    verify (F2) is now a case of §64. 218 cases, 178 of the spec.
  - `security_cms.json`, made again with the profile of v0.12: 135 cases,
    each with the lines of `Verdicts.Lines`, for each row of §29.7 and each
    item of the lists of §64 of v0.11 and v0.12: the names of the holder and
    of the issuer, the profile of the certificate, identifiers, repetitions,
    every hash and curve of the table, and the edges of the token. The two
    cases without a context, which gave the verdicts of a reader of v0.10,
    are gone.
  - `locator.json`, made again with the cases of §64 of v0.11 and v0.12: the
    first and the last address of each IPv4 block with their public
    neighbours, the IPv6 blocks, local names, characters outside RFC 3986, dot
    segments and base32 that is not a CID v1, and what the draft fixes about
    segments, case, leading zeros and the form of a CID; a locator whose
    `http` and NAT64 addresses are rejected and whose third is used; resources
    of the rest, with bytes after it, changed, cut or at another offset; data
    of the extension that a reader cannot use, a locator for another round
    among them; and plaintexts with one defect each, the bases 4094, 4070 and
    3837 completed to 4096 by a key 6 that is empty or not in its shortest
    form (change 7) among them. The bases 3837, 4070, 4094 and 4095 and those
    of the next multiple join `padding_cases`.
  - `wordkey.json`, new: the key of words of §38.1, which §64 of v0.11
    asks for and was only in the tests of `wordkey`. 45 texts and their
    words, among them each space of the list and three that are not; 20
    texts and what a writer does with them, the zero width space that it
    refuses among them; and 6 identities with their recipients, the vector
    of §38.1 and those of the next round, another `capsule_id`, another
    chain hash, round 2^53 − 1 and words that are not ASCII. For
    `datekeys-ts` and `datekeys-dart`.
  - `testdata/README.md` describes the files of v0.11 and of the draft.

## Unreleased — specification v0.11

Moves the module to the DateKeys Protocol Specification v0.11, approved by its
author on 1 October 2026 and tagged `spec-v0.11`, which adds to format 3 what
the v0.10 reserved, without changing any format: a reader of v0.10 opens these
capsules. `SpecVersion` is 0.11.

- **Area of 32 KiB.** `capsule.AreaLen` is 32768, and `LargeAreaLen`, 65536,
  which `EncryptOptions.LargeArea` lets the writer use, after the signatures
  are made, only when what they produced does not fit in 32 KiB. The fixtures of v0.10 keep their 512 bytes
  (`AreaUnit`), which a reader accepts.
- **Key of words.** Package `wordkey` derives a credential of `time_and_key`
  from six or more words that the person chooses, with PBKDF2-HMAC-SHA256 of
  600 000 iterations salted with the chain, the round and `capsule_id`
  (§38.1); the words are normalised with the tables of Unicode 18.0.0.
  `EncryptOptions.Words` adds it as one more credential, and the CLI takes
  `-words` and `-words-file` in `encrypt` and `decrypt`.
- **Author signature, `alg` 1.** `internal/ed25519strict` verifies with the
  strict profile of §29.9 and `authorkey` keeps the keys `dkauthor1…`.
  `EncryptOptions.AuthorKey` signs, inside the writer and after its checks;
  `OpenOptions.AuthorKeys` are the keys the person saved; `Verdicts` gives F2,
  F3 and F4. `capsule.PayloadCommit`, `ControlCommit`, `HeadDigest`,
  `AuthorMessage` and `SealSubject` compute what is signed.
- **Signature with certificates, `alg` 2.** `internal/cms` and `internal/der`
  read the CMS signature and the RFC 3161 token of §29.10 and §29.11 with the
  standard library only, a closed table of algorithms, and DER checked byte by
  byte. `EncryptOptions.CMSSigner` gets `AUTHOR_MESSAGE` and returns what the
  person signed outside; the writer checks it and writes nothing unless it is
  F6. The reader gives F1, F2, F5 and F6 and names the signers.
- **Time seal, `seal_type` 2.** `EncryptOptions.Sealer` asks for the token
  over `SEAL_SUBJECT`; the reader gives S1 to S5 and the authority of a valid
  seal.
- **Public note.** `EncryptOptions.PublicNote` writes `datekeys.note`;
  `Header.PublicNote` reads it; `extension.Standard` registers it.
- **Locator and envelope.** Package `locator`: the data of `datekeys.capsule`,
  the locator sealed with tlock, and the envelope split into a header and a
  rest that can hide inside another file.
- **Test data.** `format3_signed` and the vectors of its signature, and the
  fixtures `format3_signature_unsupported` and `format3_seal_unsupported`
  remade with `alg` 4294967295.
- **Review.** Three independent reviews, of the CMS reader, of the signing
  logic and writer, and of the locator and the CLI, fixed the form of the
  signature and of the TSTInfo field by field, the area decided after the
  signatures, the text of an issuer on screen and the rules of the addresses;
  spec §76 item 8 records them. Under a valid seal, `decrypt` shows an mtime
  later than the seal as an inconsistency (§29.7).
- **CLI.** `author keygen` and `author public`, `encrypt -sign`, `-note` and
  `-large-area`, and `decrypt -expect-author`. Passphrases come from a file or
  from the standard input.

## Unreleased — specification v0.10

Moves the module to the DateKeys Protocol Specification v0.10, which adds
capsule format 3, delivery 1 of its design: a capsule holds files, with
their paths, sizes, SHA-256 and modification times, a comment and a
declared author, and a security area that later versions fill with an
author signature and a time seal. `EncryptFiles` writes format 3; `Open`
and `Inspect` read the three formats, and capsules of formats 1 and 2 keep
their verdicts.

### Format 3

- `VERSION` 3 in the PRELUDE and CONTROL_CBOR of schema version 3, with the
  keys of version 2: L is the length of BODY (spec §22, §31).
- BODY (spec §29.2): a frame of three uint32, `AREA_LEN`, `SECURITY_LEN` and
  `HEAD_LEN`, the security area, 512 bytes when written and from 512 to
  65536 when read, the head and the files. A violation of the frame, L
  under 12 included, or a byte of the area after SECURITY_CBOR that is not
  zero, is `ERR_INTEGRITY` at step 17. `capsule.BodyFrame`,
  `ParseBodyFrame`, `CheckArea`.
- Security (spec §29.3, §29.7): a map of version 1 whose keys 2 and 3 are
  byte strings holding the author signature and the seal, encoded apart.
  It never fails and never decides the opening: `capsule.EvaluateSecurity`
  gives the verdicts X, F0, F1, S0, S1 and S2, and `Verdicts.Lines` the
  texts of the table, in Spanish. This version implements no `alg` and no
  `seal_type`.
- The head (spec §29.4): its type tag and version 1, then the CDDL with R1
  and R8, then, in key order, the comment and the declared author, each
  file with R2 to R6c, R10 and its layout, R7 and R9 over the tree, and the
  critical extensions. A violation of a rule of layer 4 is the new
  `ERR_HEAD_INVALID`. `capsule.Head`, `File`, `EncodeHead`, `DecodeHead`,
  and `extension.Head` for its extensions.
- Paths and texts (spec §29.5, §29.6, §29.5.1): `internal/pathrule` applies
  the rules with tables generated from 19 pinned data files, Unicode 18.0.0
  and the 15 WindowsBestFit tables, never with the Unicode functions of the
  platform; `TablesDigest` pins them. Its errors name the rule and the
  character, never echo the text, and read the same in every
  implementation. R4b and the invisibles rule of §29.6 keep ZWJ, ZWNJ,
  VS15 and VS16 to their emoji and script uses.

### Reader (spec §63 step 17, §56, §57)

- `capsule.Sink` receives the files: `Begin` with the validated head,
  `Create` for each file, `Commit` at step 18 only, and `Abort` once after
  any failure that follows `Begin`. A format 3 capsule without
  `OpenOptions.Sink` fails right after step 2 with `capsule.ErrSinkRequired`,
  a caller error without a code, before any request; a capsule of format 1
  or 2 without `dst` fails there too.
- Step 17 in its substeps: the frame and the area, security, the head, the
  files filling CONTENT, the SHA-256 of each file and the padding. A failure
  of age, or a plaintext whose length is not P, prevails; otherwise the
  first substep that fails decides, and a code other than `ERR_INTEGRITY`
  is reported only after reading PAYLOAD_AGE to its end.
- Reads of BODY grow with the bytes received, never with `AREA_LEN`,
  `HEAD_LEN` or a declared size (spec §57).
- `Opened` gains `Head`, `Verdicts`, `AreaLen` and `UnusableHeadExtensions`.

### Writer (spec §61, §62, §62.1)

- `capsule.EncryptFiles` writes format 3 from a list of `capsule.Source`,
  each read twice: first to check the paths and the texts with the rules
  of the reader, measure L with a head whose salt and SHA-256 are zero, and
  hash each file; then to write it, failing if its size or SHA-256 changed
  (rules 14 to 18). Files go in the byte order of their paths; the comment
  has its CR LF turned into LF; the mtime is kept from 1970 to 9999 and
  omitted otherwise. The head, the control and security are decoded with
  the rules of the reader before anything is written (rule 17), and the
  area is always 512 bytes with the empty security (rule 13).
- `EncryptOptions` gains `Comment`, `Author`, `HeadCritical`,
  `HeadNoncritical` and `TestVectors`. `Encrypt` writes format 2 only with
  `TestVectors`, for generators of test vectors (rule 1). `Result.Head` is
  the head written.

### CLI

- `datekeys encrypt` takes `-in` several times, files and folders; a folder
  gives its name as the first segment and is walked with `Lstat`, following
  no link, taking regular files only and leaving out `.DS_Store`,
  `Thumbs.db`, `desktop.ini`, `._*` and `__MACOSX`, which it reports. New
  `-comment`, `-author` and `-no-mtime`. A pipe is no longer accepted.
- `datekeys decrypt` writes the files of a format 3 capsule to the new
  folder `-out`: `os.Mkdir` claims it, only when there are files; the tree
  is staged in `-out/.datekeys-*` through an `os.Root`, with `O_EXCL` and
  mode 0600; the mtimes are set and each entry of the first level is moved
  into place at step 18, and any failure removes the folder. Formats 1 and
  2 still write a file.
- The presentation of spec §29.7 goes to stdout: the verdicts, the declared
  author and the comment box with their labels, the paths, and the
  verdicts again. Each line of the creator goes in pieces of at most W − 3
  columns behind `│ `, counting 2 for anything but printable ASCII, with
  its TABs expanded; W is the width of the terminal, asked through
  `syscall` on Unix and Windows, or 80. Shortcuts, `desktop.ini`, `.git`,
  programs and a leading dash get a warning.

### Errors

- `ERR_HEAD_INVALID` joins the catalogue, last (spec §69): 18 codes.

### Test data

- `datekeys.SpecVersion` is `0.10`, and every test data file says so.
- Nine fixtures of format 3 (spec §67): `format3_single`, `format3_tree`,
  `format3_comment_only`, `format3_bloque256` and
  `format3_time_and_key_portable`, written with `EncryptFiles`, and
  `format3_area_1024`, `format3_security_v2`,
  `format3_signature_unsupported` and `format3_seal_unsupported`, which only
  a generator of test vectors writes. Their records add the area, security,
  the head, the salt, the comment, the declared author, each file and the
  verdicts, and their plaintext file is BODY. The .dkk of formats 2 and 3
  join the .dkk tests (spec §68).
- New vectors: `paths.json`, `path_fold.json`, `head_schema.json` and
  `security.json`; `cbor.json` gains the control of version 3.
- The mutation corpus gains the 33 mutations of the first two lists of
  spec §64 on format 3 and the 47 of the list of format 3, three of which
  open with their verdicts; format 2 gains "format 2 time_only relabeled
  format 3", and "version changed" sets `VERSION` 4. 209 cases, 169 of the
  spec. `testkit.Splice` gives an edit for each run of changed bytes.
- The differential corpus gains two bases of format 3, one per policy:
  5110 cases, the earlier ones unchanged.
- The tests of the paths and of format 3 hold their invisible and combining
  characters as Go escapes.

## Unreleased — specification v0.9

Moves the module to the DateKeys Protocol Specification v0.9, which adds
capsule format 2 (spec §76, "Cambios normativos de la v0.9"): until the
unlock date it hides the exact length of the content and the number of
credentials. `Encrypt` writes format 2 only; `Open` and `Inspect` read both
formats, and a format 1 capsule, as v0.8.2 wrote it, keeps its verdict.

### Format 2

- `VERSION` in the PRELUDE is the capsule format, 1 or 2 (spec §22, §23):
  `capsule.Format`, `Format1`, `Format2` and `Prelude.Format`. Any other
  value is `ERR_UNSUPPORTED_VERSION` at step 2, before any request, so a
  v0.8.2 reader rejects a format 2 capsule without a network request.
- CONTROL_CBOR has the schema version of its format (spec §31). Version 2
  adds key 6, `payload_length`, L in exactly 8 bytes, big-endian, at most
  L_MAX = 2^53 − 2^46, and key 7, `padding`, 1 (bloque256) or 2
  (reforzado). A control of the other version is `ERR_UNSUPPORTED_VERSION`
  at step 14; a violation of keys 6 and 7 is `ERR_NON_CANONICAL_CBOR`.
  Without extensions a version 2 control is 103 bytes, whatever L and the
  code. `EncodeControl` and `DecodeControl` take the format.
- The payload is padded (spec §29.1): its plaintext is the content followed
  by zeros up to P = rule(L). `capsule.PaddedLength` computes both rules on
  64-bit integers, and `capsule.PayloadAgeLength` the length of PAYLOAD_AGE.
  Step 17 checks that the plaintext is exactly P bytes and its padding zero,
  or `ERR_INTEGRITY`. `Open` writes only the first L bytes to `dst`, never
  the padding (spec §56), records step 17, and reports the format, L, the
  rule and P in `Opened`.
- In format 2 INNER_ACCESS_AGE holds exactly 16 X25519 stanzas (spec §39),
  `agewrap.AccessSlots`: `agewrap.CheckAccessStanzas` and
  `agewrap.NewAccessIdentity` take the number of slots, 0 for format 1.
  Another number is `ERR_POLICY_STRUCTURE_MISMATCH` at step 12, and again at
  step 13.

### Writer rules (spec §62.1)

- `EncryptOptions.Length`, L, is required: the control is sealed before the
  content is read, and a source that delivers another number of bytes is an
  error. `EncryptOptions.Padding` chooses the rule; zero means reforzado.
- `time_and_key` takes from 1 to 16 credentials, the recipients and the
  portable key together. Each free slot gets a dummy, a fresh X25519 public
  key whose private key is dropped at once, and the 16 recipients are
  shuffled with an unbiased Fisher–Yates over `crypto/rand`.
- A recipient that is not canonical (bit 255 set, or u ≥ p) or of low order
  is rejected: `agewrap.CheckX25519Recipient`.
- Self-checks: INNER_ACCESS_AGE has 16 X25519 stanzas with distinct shares
  and the portable key opens exactly one; PAYLOAD_AGE has the length P gives
  and I_PAYLOAD opens its header.
- `Result` reports the format, L, the rule and P.

### CLI

- `datekeys encrypt` measures its input, copying a file that is not regular
  to a temporary file first, and takes `-padding reforzado|bloque256`. Its
  report and that of `decrypt` give the format and the lengths, and `decrypt`
  warns that format 1 hides neither the number of credentials nor the exact
  length. `datekeys inspect` reports the format, `"format"` in `-json`.

### Test data

- Seven format 2 fixtures (spec §67), whose records give L, the rule, P and
  the INNER_ACCESS_AGE stanza each credential opens. The five format 1
  fixtures of v0.8.2 keep their bytes as compatibility fixtures, and
  `genfixtures` never regenerates them; their records name spec 0.9, the
  format and stage 17.
- `vectors/padding.json`, new: both rules for the rows of spec §29.1,
  checked against a Padmé computed with `math/big`.
- `vectors/cbor.json`: the control vectors carry `format`, absent meaning 1,
  and gain the cases of version 2; "unknown key 6" is now "key 6, defined
  only in schema version 2".
- `vectors/mutations.json`: "version changed" writes `VERSION` 3; the 33
  mutations of the first two lists of spec §64 also run on format 2 fixtures
  ("format 2: …"), followed by the 22 of its third list and their companions
  with a `.dkk`. The new cases derive from fixtures without randomness:
  `testkit` seals them again with their known file keys and nonces
  (`reseal.go`), and reproduces each fixture byte for byte when nothing is
  edited.
- `vectors/inspect_differential.json` keeps its 1 825 cases and adds a block
  for each format 2 fixture: 4 380 cases.

## Unreleased — specification v0.8.2

Moves the module to the DateKeys Protocol Specification v0.8.2, whose
normative change closes the extension format (spec §76), refined, amended and
corrected before release (see "Specification refinements", "Specification
amendment: point canonicality" and "Specification corrections: formal
review"). Framing and schema versions do not change.

The CBOR library is replaced by a codec of the module's own, without
reflection or dependencies. Every valid object encodes to the same bytes as
before: the official vectors and fixtures are unchanged, and every error code
and inspection step of the test suite and the mutation corpus is the same.

### Specification refinements

v0.8.2 is unreleased, so these refinements amend it without a version change;
spec §76 records each with its reproducible cases. No valid object changes,
nor the verdict of any existing official vector or fixture; `dk1.json` gains
four vectors for the reading rules of §19.

- Layered error precedence (spec §69.1, with §57 and §63): within one object
  the code of the first failing layer is reported: frame, then type tag and
  schema version (keys 0 and 1), then the CBOR profile and the CDDL, then the
  fields with codes of their own in ascending key order; the rules with
  codes of their own belong to the last layer, not to the CDDL layer. Across
  objects and steps, the step order of §63 decides; step 9 now spells out its
  order (the `.dkk` as an object, then its binding to the capsule, then
  credentials, then the round time, then the request), and the errors of a
  `.dkk` come at step 9.a even when it is decoded earlier, and never under
  `time_only`. The optional inspection of steps 5, 6 and 8 and the
  `capsule_digest` check can change the code, and so could, until the
  corrections of the formal review, whether a network source verified the
  release. Steps 10, 11, 13 and 17 give the codes of the release
  verification (round, then signature) and of each identity.
- Trust model (spec §55.1): who can write PUBLIC_HEADER, CONTROL_CBOR,
  PAYLOAD_AGE and the `.dkk` body, from which step each is bound and by what,
  and what none of them proves. Spec §72: an extension with security-relevant
  claims lives in CONTROL_CBOR or is signed by a signature extension; the
  data of `.dkk` extensions is advisory for its holder only.
- One ordering rule for extension arrays (spec §31, §54, CDDL): strictly
  ascending unsigned bytewise order of the UTF-8 bytes of `extension_id`, a
  proper prefix first, never UTF-16 code units or a locale collation.
- Rules that only `testdata/README.md` stated are now normative text: the age
  header grammar of C2SP and its codes (§28.1, §36), the round-time bound of
  9999-12-31T23:59:59Z and pre-genesis instants (§15), the reading rules of
  `dk1_` (§19: either Base64 alphabet but no CR or LF, one JSON object in
  valid UTF-8 without a byte order mark, numbers by their exact decimal
  value), lengths
  of at least 1 (§22, §23, §40, §57), the comparison of the tlock stanza
  arguments (§35, §63 step 8), the Provider Profile rules and the chain-hash
  formula (§12.1, with `period` at most 2^32 − 1 and the name alphabets
  normative; the name lengths stay implementation limits), `extension_id` of
  at least one byte (§31), a repeated X25519 ephemeral share in
  INNER_ACCESS_AGE (§36), and the implementation limits of the reference
  (§74).

These error codes of the reference change, for inputs that no existing
official vector holds:

- `capsule.Open` checks an offered `.dkk` as an object before binding it to
  the capsule: `access_type` and `access_material`, then its critical
  extensions, then `capsule_id`, then `capsule_digest`. A `.dkk` for another
  capsule with an unknown critical extension is now
  `ERR_EXTENSION_CRITICAL_UNKNOWN` (was `ERR_ACCESS_INVALID`), and one with
  an unsupported `access_type` and an unknown critical extension is
  `ERR_ACCESS_INVALID` (was `ERR_EXTENSION_CRITICAL_UNKNOWN`).
- `accesskey.Decode` rejects a `BODY_LEN` of 0 with `ERR_INTEGRITY`, like a
  `PUBLIC_HEADER_LEN` of 0; it was `ERR_NON_CANONICAL_CBOR`, the empty body
  failing to decode.
- `datekey.Parse` rejects CR and LF in a `dk1_` string with
  `ERR_DATEKEY_INVALID`; the Go Base64 decoders skipped them, and the
  result was `ERR_DATEKEY_NON_CANONICAL`. It also rejects invalid UTF-8 in
  the JSON at step 2 (see "Fixed").
- `datekeys decrypt -dkk` hands the `.dkk` to `capsule.Open` still encoded,
  through the new `OpenOptions.AccessKeyFile`, so its decoding errors come at
  step 9 of a `time_and_key` capsule, after any failure of steps 1 to 8, and
  a `time_only` capsule ignores it; the CLI used to fail on it before
  reading the capsule.
- `capsule.Open` ignores nil entries of `OpenOptions.Identities`: with no
  other credential, the result is `ERR_ACCESS_REQUIRED` at step 9, before
  the clock is consulted, instead of `ERR_RELEASE_UNAVAILABLE` or a release
  request.
- `agewrap.AccessIdentity` tries every identity on every stanza: an identity
  that unwraps two INNER_ACCESS_AGE stanzas is
  `ERR_POLICY_STRUCTURE_MISMATCH` even when another identity unwraps exactly
  one, whatever their order; the file used to open when the other came
  first.
- `profile.NewRegistry` encodes and decodes each profile before comparing its
  `profile_hash`, so a profile gets the code `profile.Decode` reports: a
  `period` of 86401 s is `ERR_NON_CANONICAL_CBOR`, as in `Decode`, not
  `ERR_UNKNOWN_PROFILE`. `Profile.Validate` checks the schema rules a value
  can break (a period that is not a whole number of seconds in 1..86400, a
  genesis time outside 0..2^53−1, a name that is not valid UTF-8) first,
  with `ERR_NON_CANONICAL_CBOR`.

Tests: `capsule.TestPrecedenceWithinPublicHeader`, `TestPrecedenceAcrossSteps`,
`TestFrameLengthLowerBounds`, `TestMalformedAgeHeaders`,
`TestTlockStanzaArgumentComparison`, `TestAccessKeyCheckOrder`,
`TestAccessKeyFileAtStep9`, `TestControlCriticalBeforeHeaderBinding` and
`TestTrustModel`; `accesskey.TestDecodePrecedence`;
`agewrap.TestAccessIdentityStrictness`; `datekey.TestReadingRules`;
`profile.TestDecodePrecedence` (a G1 point outside the prime-order
subgroup), `TestPinPathMatchesDecode` and `TestChainHashFormula`;
`provider.TestVerifyRejects`; `extension.TestOrderIsUnsignedBytewise`;
`cmd/datekeys.TestDecryptAccessKeyOrder`.

### Specification amendment: point canonicality

An amendment of the unreleased v0.8.2, recorded with its case in spec §76.
The new §12.2 defines the canonical encoding of a BLS12-381 point, the
compressed form of drand: the compression flag set, the infinity flag only for
the point at infinity with every other bit zero, the sort flag for the
lexicographically largest y, big-endian coordinates below p (c1 then c0 in
G2) and a point of the prime-order subgroup. A decoder rejects every other
string, among them x + p and an identity with a payload. The Provider Profile
public key (§12.1), the release signature (§63 step 10) and the U of the
tlock stanza (§63 step 11) are canonical and never the point at infinity, and
step 11 defines the stanza body, `U || V || W` with |V| = |W| = 16 (128 bytes
for Quicknet), and the IBE check r·G == U. The case, from the second
implementation: `tlock-js` on `@noble/curves` 1.9.7 accepted U re-encoded as
c0 + p and a signature re-encoded as x + p and returned the same file key,
where the reference rejects both.

No error code or step of the reference changes: its decoder,
`kilic/bls12-381` through drand, already rejected those encodings. A U at
infinity, which only the IBE check used to reject, is refused before
decryption, with the same `ERR_INTEGRITY`. Spec §64 gains ten mutations,
exported in `mutations.json`: U with c0 + p, U at infinity, U with the
infinity flag and a payload, and tlock stanza bodies of 127 and 129 bytes,
each with a valid header MAC (`ERR_INTEGRITY`, step 11); a release signature
with x + p, at infinity, with the infinity flag and a payload, negated, and
negated together with U with c0 + p (`ERR_RELEASE_INVALID`, step 10). The
x + p case is a capsule for round 1004, whose published signature `testkit`
now knows (`testkit.XPlusPRound`): no fixture round has an x below
2^381 − p. No fixture and no other vector changes.

Tests: `profile.TestDrandPointDecodersAreCanonical`, which fails if a
dependency update makes the decoders of drand lenient, and
`TestPublicKeyEncodingIsCanonical`; new cases in `provider.TestVerifyRejects`,
`agewrap.TestTimeIdentityStrictness` and `TestTimeIdentityRelease`;
`capsule.TestPointMutationsChangeOnlyTheEncoding`, which checks that each
point mutation differs from a capsule that opens only in one encoding;
`internal/testkit.TestPointReencodings`.

### Specification corrections: formal review

Corrections of the unreleased v0.8.2 from the formal review of the
specification, recorded with their cases in spec §76 ("Correcciones de la
revisión formal"). Three change normative rules:

- A source that fetches releases over a network (a relay, the Release API or
  a cache) must verify every response with the rules of step 10 and discard
  the one that fails; when none passes, the code is `ERR_RELEASE_UNAVAILABLE`
  at step 9 (spec §63). The codes of step 10 are those of a release supplied
  directly, as in the official vectors. `provider/drand.Client` already did
  this; the contract of `provider.ReleaseSource` and the documentation of
  `capsule.OpenOptions.Source` now say so.
- Each registered extension declares the objects (PUBLIC_HEADER,
  CONTROL_CBOR, `.dkk`) and the arrays where it may appear, and a known
  extension elsewhere is treated as unknown there (spec §31, §54, §72): a
  critical one is `ERR_EXTENSION_CRITICAL_UNKNOWN`, a noncritical one is
  ignored. The new optional interface `extension.Placement` of a `Registry`
  tells where each extension is registered; `capsule.Inspect` and
  `capsule.Open` check the arrays of PUBLIC_HEADER (step 4), of the `.dkk`
  (step 9.a) and of CONTROL_CBOR (step 14) with the new
  `extension.CheckCriticalIn` and `extension.CheckNoncriticalIn`. A
  `Registry` that does not implement it behaves as before.
- H2, H3 and H4 of step 11 are those of drand/kyber `encrypt/ibe`, and H2
  hashes the element of GT in the order of `kilic/bls12-381`: c1 before c0 at
  every level of the tower, each coordinate of Fp in 48 bytes big-endian. The
  new vector file `testdata/vectors/tlock_ibe.json` freezes H2(e(G1, G2)) =
  `cb87319f24560b5231579a09ad79f12e`; the order of `Fp12.toBytes` of noble,
  c0 first, gives `0118eea9d5971745f71e3c94926f1717` and another file key.

The other findings are editorial: §27 defers the authenticity of
PUBLIC_HEADER to the trust model (§55.1) and says that the age header MAC
protects only against whoever does not know the file key; step 5 separates
the mandatory read of SEALED_CONTROL from the optional inspection of its age
header; step 15 names `ERR_HEADER_BINDING`; §21 makes the 16 CSPRNG bytes of
`capsule_id` a MUST; §77 gains drand/kyber, RFC 8259, RFC 8610 and RFC 4648;
§76 retitles its section "Cambios normativos de la v0.8.2" and corrects its
record: `dk1.json` gained four vectors, not three, and two cases of the
refinements quoted texts that no earlier version contained.

No error code of the reference changes, nor any fixture or existing vector.
`github.com/drand/kyber-bls12381`, already an indirect dependency through
drand and tlock, becomes a direct requirement: `internal/testkit` and a test
compute the pairing with it.

Tests: `capsule.TestReleaseFromANetworkSource` (a relay whose only answer is
a release of another round, or a negated signature, gives
`ERR_RELEASE_UNAVAILABLE` at step 9 through `provider/drand.Client`, and the
same release supplied directly the code of step 10);
`capsule.TestExtensionPlacement` (an extension registered for CONTROL_CBOR
only is unknown in PUBLIC_HEADER and in a `.dkk`, one registered as
noncritical only is unknown in a critical array, and a noncritical copy
outside its registration is ignored); `extension.TestPlacement`;
`agewrap.TestTlockH2Vector` (the frozen vector, and step 11 recomputed with
H2 and H4 against the file key that tlock unwraps).

A second round of the same review confirmed these corrections and asked for
three more, recorded in spec §76 as corrections 4 to 6:

- An encoder must not write a registered extension in an object or array it
  is not registered for, and a reader that ignores a noncritical one for that
  reason must not interpret its data (spec §54, §72). A writer could seal in
  the critical extensions of CONTROL_CBOR one registered there as
  noncritical only, which a reader that knows the registration rejects at
  step 14, after the unlock. `capsule.Encrypt` and `accesskey.Encode` take
  no `Registry`: they write the extensions they are given, and the
  application, which knows the registration, applies the rule.
- §17 and §51 give the codes of step 10 only for a release supplied
  directly, as step 10 does; one that a network source fetches and that
  breaks those rules is discarded at step 9.
- Whatever the failure of the release source, step 9 reports
  `ERR_RELEASE_UNAVAILABLE` and no other code (spec §63).

The other findings are editorial: §28.1 says that steps 5 and 6 parse only
the age header, since step 5 reads all of SEALED_CONTROL; step 9 has a single
arrow to `ERR_RELEASE_UNAVAILABLE`; §76 rewords two introductions.
`testdata/README.md` says where the extensions of the mutation corpus are
registered: in both arrays of every object.

These error results of the reference change, for inputs that no official
vector holds:

- `capsule.Open` reports any error of `OpenOptions.Source` at step 9 with
  `ERR_RELEASE_UNAVAILABLE` as its only code. An error that carries another
  normative code, or none, keeps only its text: a caller's source that
  failed with `ERR_RELEASE_INVALID` gave `ERR_RELEASE_INVALID` at step 9. An
  error that wraps `ERR_RELEASE_UNAVAILABLE` alone is returned as it is, with
  its other causes, such as a context error.
- The error of `provider/drand.Client.Fetch` wraps `ERR_RELEASE_UNAVAILABLE`
  and no other normative error, as `errors.go` promises for every error of
  the module. The failure of each relay is kept in its text only, where
  `errors.Join` wrapped it: a relay that answered with a release of another
  round also made the error match `ERR_ROUND_MISMATCH`. A context that ended
  is still wrapped, whether `Fetch` sees it end before or after the relays
  fail because of it.

Tests: `capsule.TestReleaseSourceErrorsAtStep9` (a source error with another
code, two codes or none gives `ERR_RELEASE_UNAVAILABLE` alone at step 9 and
keeps its text); `capsule.TestExtensionPlacement`, which now also covers the
noncritical array of a `.dkk`, so that checking it with the object-blind
`extension.CheckNoncritical` fails; `provider/drand.TestFetchErrorHasOneCode`
and `TestUnavailabilityAndCancellation` (a canceled context and a deadline
during the request); `datekeys.TestCode`.

### Breaking changes

- `extension.New(id, version, data []byte)` takes the opaque data bytes instead
  of a value that it encoded as CBOR, and rejects nil or empty data. An
  extension without data is the literal `extension.Extension{ID, Version}`,
  which omits key 2.
- Extension data (key 2) must be a byte string of at least one byte. Any other
  CBOR type, `null` or `h''` at key 2 is now `ERR_NON_CANONICAL_CBOR`, so a
  v0.8.1 object with such data no longer decodes. The base protocol never
  decodes the content (§54).
- `codec.Valid` and its fuzz target `codec.FuzzValid` are removed: nothing
  decodes extension data any more.
- Package `codec` is rewritten without reflection, struct tags or dependencies
  (§58). Removed: `Marshal`, the reflection-based `Unmarshal(data, v)` and
  `Peek(data, v)`, and `MaxNestedLevels`, `MaxArrayElements` and
  `MaxMapPairs`. Each schema now writes its encoding with a `codec.Encoder`
  (`Map`, `Array`, `Uint`, `Bstr`, `Text`, `Out`, with a sticky first error,
  and `Fail`, which records an error of the schema so that `Out` never returns
  bytes its decoder rejects) and reads it with a strict `codec.Decoder`
  (`NewDecoder`, `Map`, `Key`, `EndMap`, `Array`, `Uint`, `Bstr`, `Text`,
  `Done`). `codec.Unmarshal(in, decode, encode)` runs the decoder of a schema
  and requires that its re-encoding reproduces the input; `codec.Peek(in)`
  returns the type tag, of at most `codec.MaxTypeTagLen` (64) bytes, and the
  schema version; `codec.Walk(in, maxDepth, maxLen)` checks that bytes are one
  item of the §58 profile, for vectors, fuzzing and diagnostics.
  `CheckSchema` and `MaxSafeUint` keep their names.
- `extension.Wire` and its `UnmarshalCBOR` are removed. `extension.Encode`
  becomes `extension.Canonical`, which returns the validated array in
  canonical order as `[]Extension`; `extension.Decode([]Wire)` becomes
  `extension.DecodeArray(*codec.Decoder)`, which reads and validates one array
  and rejects more than 64 entries from the array head, before reading any;
  `extension.EncodeArray(*codec.Encoder, []Extension)` writes one, and
  records in the Encoder, instead of writing it, an array that `DecodeArray`
  would reject.
- `codec.CheckSchema` reads keys 0 and 1 only, and nothing after them (§70):
  the map head, key 0, a type tag of at most 64 bytes, key 1 and the version
  must be in the profile, each head in its shortest form, and the version at
  most 2^53−1. A schema version other than the expected one read that way is
  `ERR_UNSUPPORTED_VERSION` whatever follows it; it was
  `ERR_NON_CANONICAL_CBOR` when the rest of the object was malformed. Every
  other form of the version is now `ERR_NON_CANONICAL_CBOR`, where it was
  `ERR_UNSUPPORTED_VERSION` whenever the value read was not the expected one:
  a missing version, `null` or `undefined`, a version not in its shortest
  form, a version above 2^53−1 (up to 2^64−1), a version that is not the
  second key (placed before key 0 or after another key), and a version
  behind a map head or a type tag head not in its shortest form. `true` and
  `false` were already `ERR_NON_CANONICAL_CBOR`.
- `capsule.DecodeHeader` checks every CDDL rule of PUBLIC_HEADER, including
  `access_policy`, the extension arrays and the cross-array rule, before it
  parses the DateKey (§57, §63 step 4). A header that breaks both reports
  `ERR_NON_CANONICAL_CBOR` where it reported `ERR_DATEKEY_INVALID` or
  `ERR_DATEKEY_NON_CANONICAL`; a header with one fault keeps its code.
- `profile.Profile.CanonicalCBOR` and `Hash` refuse a `profile_id`,
  `provider`, `network` or `scheme` that is not valid UTF-8, with
  `ERR_NON_CANONICAL_CBOR`: they wrote it as an invalid text string. Every
  encoder refuses such text.
- At most 64 extensions per array and `extension_version` at most 2^32−1, on
  encode and decode (`ERR_NON_CANONICAL_CBOR`). An `extension_id` appears at
  most once per object, and arrays are ordered by `extension_id` only.
- Provider Profile: `genesis_time` is an unsigned integer, and `period` and
  `genesis_time` are at most 2^53−1; a negative or larger value is
  `ERR_NON_CANONICAL_CBOR`.
- `null` in a byte-string field is `ERR_NON_CANONICAL_CBOR` (for example a
  `null` `access_material` was `ERR_ACCESS_INVALID`): nil byte strings, arrays
  and maps now encode as empty ones, never as `null`.
- `capsule.EncodeHeader` and `capsule.DecodeHeader` enforce the 1 MiB
  PUBLIC_HEADER limit and `accesskey.DecodeBody` the 16 MiB BODY limit. Every
  frame-limit refusal, including those of `accesskey.MarshalBody` and of
  `capsule.Encrypt` for SEALED_CONTROL, now wraps `ERR_INTEGRITY` (§57).
- `profile.Profile.CanonicalCBOR` refuses a `period` or `genesis_time` outside
  the schema with `ERR_NON_CANONICAL_CBOR`, as `profile.Decode` does.
- The official fixture `time_only_extensions` is regenerated: its header data
  is the raw UTF-8 bytes of "public label" and its control data is
  `{0: 7, 1: "sealed"}` (`a2000701667365616c6564`). Every other `.dkc` and
  `.dkk` keeps its bytes; the fixture and vector metadata name spec 0.8.2.

### Added

- `datekeys.SpecVersion` (`0.8.2`), the specification the module implements,
  and `datekeys.Version()`, the version of the module as the go command
  recorded it: a tag, the pseudo-version of the commit of a checkout build,
  or `(devel)`. `datekeys version` prints both and the Go toolchain.
- `capsule.OpenOptions.AccessKeyFile`, a `.dkk` still encoded, which `Open`
  decodes at step 9.a and only for a `time_and_key` capsule (§63, §69.1).
- `ErrExtensionDataInvalid` (`ERR_EXTENSION_DATA_INVALID`, §69).
- `extension.Placement`, an optional interface of a `Registry` that tells in
  which objects and arrays each extension is registered, with
  `extension.Object` (`PublicHeader`, `Control`, `AccessKey`),
  `extension.Array` (`Critical`, `Noncritical`), `extension.KnownIn`,
  `extension.CheckCriticalIn` and `extension.CheckNoncriticalIn` (§54, §72).
  `CheckCritical` and `CheckNoncritical`, which do not know the object, keep
  their behaviour and consult no `Placement`.
- `testdata/vectors/tlock_ibe.json`, the H2 vector of §63 step 11, generated
  by `internal/testkit.IBEVectors` and documented in `testdata/README.md`.
- `extension.DataValidator`, an optional interface of a `Registry` that
  validates the data of the extensions it knows: a known critical extension
  with invalid data fails with `ErrExtensionDataInvalid` (§63 steps 4 and 14,
  and the `.dkk` check); a known noncritical one is reported in
  `capsule.Inspection.UnusableExtensions`, `capsule.Opened.UnusableControlExtensions`
  or `capsule.Opened.UnusableAccessKeyExtensions` (`extension.CheckNoncritical`,
  `extension.Unusable`) and does not fail.
- Encoder self-checks: `capsule.Encrypt` decodes its PUBLIC_HEADER and
  CONTROL_CBOR, and `accesskey.MarshalBody` its body, with the readers'
  decoders before sealing or writing (§72).
- `extension.MaxExtensions`, `MaxVersion`, `MaxDataLen`, `codec.MaxSafeUint`
  and `codec.MaxTypeTagLen`.
- The `.dkk` fixture `time_and_key_portable_extension`, which carries a
  noncritical extension with data (§68).
- `genfixtures -only NAME[,NAME...]` regenerates the named fixtures only.
- Tests: the three new §64 mutations (data that is not a byte string, `h''`
  data, 65 extensions), regression tests for the cases of §76, conformance
  checks on the exact data bytes, and the fuzz target
  `capsule.FuzzEncodeImpliesDecode` (header, control and `.dkk`).
- The mutation corpus moves from `capsule/mutation_test.go` to
  `internal/testkit.Mutations`, shared by the test and by `genfixtures`. Its
  third-party X25519 identity is now fixed (`testkit.Stranger`), and every
  release source answers with one recorded release, as the exported corpus
  describes it. The CLI's inspect view moves to `internal/inspectview`, which
  `genfixtures` uses to freeze the outputs.
- `internal/cbortest`, an encoder and decoder of generic CBOR values written
  independently of `codec`: tests build with it inputs outside the profile
  and check `codec` against it.
- Tests of the map structure of every schema (key order, required and unknown
  keys, entry counts) and of the codec, whose statement coverage is 100 %.
- `access_policy` values whose low byte is 0 or 1 (256, 257, 65536, 2^32,
  2^53−256…) are tested as undefined, as `FuzzDecodeHeader` seeds and as two
  mutations with a consistent `header_binding` (`testkit.Build.RawPolicy`).
- Tests `extension.TestEncodeArrayRejects`,
  `accesskey.TestEncodeAndDecodeLeaveNoStaleMaterial`,
  `accesskey.TestDecodeShortBodyAllocatesLittle` and `capsule.TestDecryptAll`.
- Shared test data for a second implementation, generated by `genfixtures`,
  regenerated by the gate and documented in `testdata/README.md`:
  - `testdata/vectors/cbor.json`: 36 accepted and 67 rejected items of the
    §58 profile, walked with `codec.Walk` (integers above 2^53−1 as decimal
    strings), and 135 schema vectors: minimal valid object, unknown key,
    missing key, wrong type, size and range for the Provider Profile,
    PUBLIC_HEADER, CONTROL_CBOR, the `.dkk` body, `verification_metadata` and
    extensions (data `40` and `5801xx`, data of every other type, 64 and 65
    extensions, a leading BOM, the U+FF61/U+10000 order, `extension_version`
    2^32−1 and 2^32); schema versions 2^53 and 2^64−1 and the order of type
    tag and version; and the Provider Profile validation (names, public key,
    `genesis_time`, drand scheme, the chain-hash self-check with its formula,
    the `period` limit), each vector keeping the chain hash consistent unless
    it tests the self-check.
  - `testdata/vectors/mutations.json`: the mutation corpus as frozen data, 65
    cases (the 33 of §64 first), each a `.dkc` given as edits of a fixture and
    what the reader is given (`.dkk`, identities, the recorded release, clock,
    registry, known extensions), with the expected error and step. The 17
    capsules built with age randomness are kept from the committed file;
    `genfixtures -only mutations` rebuilds them.
  - `testdata/vectors/inspect_differential.json`: 1825 deterministic mutations
    of the five `.dkc` fixtures (bit flips, byte changes, truncations,
    insertions, deletions, length fields, CBOR-aware header edits, DateKey
    edits, age header edits) with the verdict of steps 1 to 8.
  - `testdata/fixtures/<name>.inspect.json`: the exact output of
    `datekeys inspect -json -in <name>.dkc` for each official capsule.
  - `testdata/README.md` points to the sections of the specification that
    decide each verdict (§12.1, §15, §19, §22, §23, §28.1, §63, §69.1, §74),
    which the refinements above moved into normative text.
- `spec/datekeys.cddl` marks the one-day `period` limit of the Provider
  Profile, which `profile.Decode` already applied, as an implementation limit
  of the reference (§57, §74); spec §11 allows up to 2^53−1.
- Tests that replay them: `codec.TestSharedVectors`,
  `internal/testkit.TestSchemaVectors`, `capsule.TestExportedMutationCorpus`,
  `capsule.TestInspectDifferentialCorpus` and
  `cmd/datekeys.TestInspectJSONGoldens`.
- Fuzz targets `codec.FuzzDecoder` (the Decoder primitives),
  `codec.FuzzWalk` (against the independent decoder), `codec.FuzzPeek`,
  `codec.FuzzEncodeImpliesWalk` and `extension.FuzzDecodeArray`, run by
  `scripts/fuzz.sh`; `codec.FuzzUnmarshal` now fuzzes a hand-written schema.

### Removed

- The dependencies `github.com/fxamacker/cbor/v2` and
  `github.com/x448/float16`. `go.mod` requires nothing new.

### Fixed

- Error messages no longer copy the text of an error of age, tlock, kyber,
  drand or kyber-bls12381. When the IBE check of a tlock stanza failed,
  kyber's error carried the candidate plaintext and r, and
  `agewrap.TimeIdentity` copied it into its error, and so into the error of
  `capsule.Open` and the details of `Inspection.Checks`: a third party who
  edited W learned FK_TIME from the message. Each such failure now has a
  fixed message with its normative error and a reason of its own: the length
  of the tlock stanza body, the encoding of U, U at infinity or the IBE check;
  the header or the STREAM of an age file; a malformed X25519 stanza. A
  failure of the writer of the plaintext at step 17 keeps its own text.
  `capsule.TestTlockFailureDiagnosticsCarryNoSecrets`,
  `TestPlaintextWriterFailureKeepsItsText`.
- `datekey.Parse` rejects invalid UTF-8 in the `dk1_` JSON at step 2, as
  §19 requires. `encoding/json` replaced it with U+FFFD, so a member that a
  repeated name overwrites passed steps 2 and 3 and ended as
  `ERR_DATEKEY_NON_CANONICAL` instead of `ERR_DATEKEY_INVALID`. Found by the
  second implementation's differential; new `dk1.json` vector.
- `extension.CheckDisjoint` is a linear merge of the two sorted arrays; a
  PUBLIC_HEADER with 40 000 + 40 000 extensions took 8.3 s in the pairwise
  check (§76, case 6).
- Control data made of 14 or 15 nested arrays was sealed by `Encrypt` and
  rejected by `Open` at step 14, after the unlock (§76, case 5).
- Header data `{NaN: 0, NaN: 1}` gave a nondeterministic verdict (§76, case 3).
- `extension.New(id, v, nil)` wrote `null` as data (§76, case 2).
- `codec.Unmarshal` wipes its re-encoding, which after the new self-checks
  held a copy of I_PAYLOAD or `access_material`, and `accesskey.DecodeBody`
  wipes the material on its error paths. The `codec.Encoder` also wipes every
  buffer it outgrows, and `codec.Unmarshal` sizes its re-encoding for the
  input; the former library's internal buffers could keep a copy.
- `accesskey.Encode` wipes the body it wrote, and `accesskey.Decode` reads the
  body into a buffer that grows with the data read, wiping every buffer it
  outgrows, and wipes the body once decoded or on error: both left copies of
  `access_material` behind. The in-memory age decryption of SEALED_CONTROL and
  INNER_ACCESS_AGE in `capsule.Open` reads the plaintext into one buffer of
  the ciphertext's size instead of a growing one, so that no outgrown buffer
  keeps a copy of I_PAYLOAD. Buffers internal to `filippo.io/age` and copies
  made by the Go runtime stay out of reach (SECURITY.md).

## Unreleased — v0.1.0

First implementation of the DateKeys Protocol Specification v0.8.1.

### Added

- `datekey`: local date → round resolution at full precision (§15), canonical
  `dk1_` encoding and strict parsing (§18, §19).
- `profile`: Provider Profile Deterministic CBOR and `profile_hash` (§11), the
  pinned Quicknet profile with its chain-hash self-check (§12), and pinned
  registries (§13).
- `provider`: release sources and local BLS verification (§51);
  `provider/drand`: racing public relays, verifying every answer (§48, §49, §52).
- `codec`: Deterministic CBOR with a re-encoding canonicality check (§58, §58.1).
- `extension`: the generic extension mechanism (§54).
- `agewrap`: strict tlock and X25519 age identities that enforce the stanza
  rules (§27, §29, §32, §33, §35), and a secret-free header probe.
- `capsule`: `.dkc` framing, `Encrypt` for `time_only` and `time_and_key`
  (§61, §62), `Inspect` (§63 steps 1–8) and `Open` (§63 steps 9–18).
- `accesskey`: `.dkk` encoding and decoding (§40–§44).
- `cmd/datekeys`: `encrypt`, `decrypt`, `inspect`, `datekey resolve`,
  `profile hash`, with atomic, non-overwriting outputs.
- Official vectors (§65, §66), `.dkc`/`.dkk` fixtures (§67, §68), the mutation
  corpus (§64), fuzz targets for every parser, interoperability tests with the
  official `age` and `tle` CLIs, and live Quicknet integration tests.
- `spec/datekeys.cddl` and `docs/traceability.md`.