package wire

1.2.0

Added

• Support wasm_of_ocaml (31-bit int) and js_of_ocaml (32-bit int). Codecs no longer silently truncate values that exceed the target's native integer range (#232, @samoht)
• Add Wire.Expr.lsr64, an int64 logical shift right by a constant amount, for isolating the high bits of a full-width field inside a constraint. It projects to a 3D refinement EverParse verifies (#232, @samoht)
• Add Wire.Codec.field_readers and Wire.Everparse.Raw.int_slots. field_readers is the decoder's own reader for every named int-valued field, keyed by name, so a caller holding a type-erased codec can read a field without a field handle; int_slots is the byte slot of every named whole-byte integer field (#318, @samoht)
• Add Wire.Everparse.Raw.field_seeds, exposing constrained whole-byte integer values, and seed Wire_3d.generate_corpus from them and from accepted inputs, so a generated corpus reaches values uniform bytes never would. A one-sided corpus, which cannot tell a validator from a constant answer, is now refused (#314, @samoht)
• Add Wire.equal_error_kind, structural equality on error kinds ignoring location, and Wire.Everparse.equal_field_action_form (#299, #302, @samoht)
• Document how to make a whole field group conditional with Field.optional over a sub-codec, keeping the group's byte-length prefix and dependent Field.repeat local to the group while later fields stay in the parent codec (#236, @samoht)

Changed

• Breaking: A fixed-width integer field decodes to a carrier whose range is the field's, so a value it cannot hold is refused where it is built instead of being masked into a legal-looking one on the way to the wire: uint8 to Wire.UInt8.t, int8 to SInt8.t, uint16 and uint16be to UInt16.t, int16 and int16be to SInt16.t, uint32 and uint32be to UInt32.t (was Optint.t), int32 and int32be to SInt32.t, uint64 and uint64be to UInt64.t (was int64). The 8- and 16-bit carriers are a private int, read with to_int and built with the range-checked v; the wider ones are abstract, read with to_int32 / to_int64 and built with of_int32 / of_int64 or the range-checked of_int. SInt32 and UInt32 are deliberately not interchangeable: they share a representation, so a value read with the wrong signedness would be a silent misparse rather than a type error (#321, #322, #323, #330, #333, #335, #336, @samoht)
• Breaking: Wire.uint decodes to Optint.Int63.t rather than a native int, spelled as that type rather than the Wire.Private alias: a 7-byte value needs 56 bits, which used to truncate on a narrow-int target. Read it with Optint.Int63.to_int / to_int64 (#232, #270, @samoht)
• Integer-valued combinators no longer insist on a native int base. Wire.enum, enum_open, variants, lookup and bit sit over uint32, int32 or uint64 as readily as over uint8, and project to EverParse the same way; Wire.rest_bytes takes a size parameter of any integer type, and Wire.Field.int64 a field of any type carrying a 64-bit slot (#321, #323, #328, @samoht)
• Breaking: Require OCaml 5.3, up from 5.2 (#320, @samoht)
• Breaking: Codec.encode and Wire.to_string raise Invalid_argument on a value their own decoder rejects, rather than writing bytes that fail to read back: an unlisted value in a closed enum, a non-zero byte in an all_zeros padding field, a byte_array_where byte that fails its ~per_byte refinement, and a record whose where clause or field ~constraint_ does not hold. Encoding a well-formed value is unchanged (#263, @samoht)
• Breaking: Encoding an integer a field cannot hold raises Invalid_argument instead of dropping the bits that do not fit, so int8 refuses 200 rather than writing it back as -56, and bits ~width and uint ~size accept exactly their declared width. A masked value is itself legal at that width, so nothing downstream could tell it from a value meant that way. Applies to Wire.to_string, Codec.encode and Codec.set alike (#275, @samoht)
• Breaking: Encoding a byte_array, byte_array_where or byte_slice whose length differs from its declared ~size raises Invalid_argument. It used to truncate a longer value, zero-pad a shorter one, or, through the streaming writer, desynchronise every following field, so the bytes on the wire could differ from the value a caller had signed, hashed or length-prefixed. For a short value in a fixed-size region use zeroterm_at_most, which is NUL-terminated and round-trips (#259, @samoht)
• Require bytesrw >= 0.4.0 (#297, @samoht)
• Codec.load_word reads the bitfield base word as an Optint.t (was a native int) and Codec.extract takes it, so the word-at-a-time batch API is exact for 32-bit bases on every platform; on a 64-bit host the word stays an unboxed int (#232, @samoht)
• Change the exception raised for a description no decoder or encoder can handle, such as a casetype value no case projects or an unresolved type reference, from Failure to the Invalid_argument documented in wire.mli. The buffer is untouched on those paths, so only the exception type changes (#291, @samoht)
• Change the zeroterm and zeroterm_at_most encode errors to report as Wire.encode on every path. Codec.encode reported the same two faults, a NUL in the value and a value too long for its region, under its own name (#269, @samoht)
• Change the generated .3d C struct tags to Wire<Name> rather than _<Name>. Public typedef names are unchanged, and standalone entrypoints follow as <Base>CheckWire<Name> (#235, @samoht)

Fixed

• A codec that embeds a sub-codec now projects to a schema EverParse accepts whatever that sub-codec holds. The types a sub-codec's own fields name were declared after it rather than before, and a refined byte span reached only through a sub-codec was never declared at all, so EverParse rejected the schema with a "not found" naming a type the codec does use (#344, @samoht)
• Generating C for a codec with a long name no longer fails with Invalid_argument "List.iter2". Past a certain length the generated callback declarations wrap onto a second line, which the field-plug generator read as no declarations at all; both layouts are read now, and a header that still cannot be read is named in the error rather than surfacing as a crash somewhere else (#344, @samoht)
• A value with no fixed wire size read through Wire.of_reader no longer costs time and memory quadratic in its length. The reader rebuilt the whole accumulated buffer and re-parsed it from offset zero after every slice, so a 64 KiB zeroterm arriving one byte at a time took 1.7 s where it now takes a millisecond, at the top level and inside a codec alike (#325, #331, @samoht)
• Codec.validate allocates a constant amount whatever the sender sends, and nothing at all for a codec with no parameters. It used to run the field readers for their checks and drop what they built, so a refined span, a zeroterm_at_most, an all_zeros, a repeat's elements and a closed enum's named cases each cost memory in proportion to lengths and counts the sender chooses. Sub-codecs reached through an array, a nested region or a casetype case body are checked in place too, and Codec.decode no longer keeps a second copy of a span it returns (#286, #289, #290, #292, #293, #294, #295, #329, @samoht)
• A fixed-offset Codec.get or Codec.set with no parameters, and Wire.of_string / Wire.of_bytes on a struct type, no longer allocate per call: direct scalars represented as immediate OCaml values, bitfields, enums and immediate-valued maps measure zero words on a non-Flambda release build, and the validator cache lookup no longer builds a closure and an option. Boxed results such as int64, and allocations performed by a map callback itself, remain visible to callers. The same accessors are also faster than before (#239, #269, @samoht)
• A fresh domain's first decode or validate no longer costs about a word per compiled validator the program ever built, which reached roughly 1 MB at 50,000 validators and was charged even to domains that never decode (#320, @samoht)
• Codec.decode and Codec.validate no longer accept bytes that Wire.of_string and the EverParse validator built from the same schema reject. A sub-codec field's constraints, where, closed-enum membership, ~per_byte refinements and actions were skipped, a byte_array_where's ~per_byte ran on the direct parser alone, and a Field.repeat whose byte budget does not split into whole elements passed (#268, #305, #311, @samoht)
• Codec.validate no longer raises Invalid_argument where a parse error was due. It reaches a field at whatever offset the layout computes and those reads were unguarded, so an overlong variable-size field pushing its successors past the end of the buffer crashed the documented gate for untrusted input. Same for Wire.of_bytes and Wire.of_string on a struct, an optional gated on a runtime expression, and a uint of runtime width. All now report end of input at the field's own offset (#271, #276, @samoht)
• Codec.get and Codec.set no longer read or write past the field they name. A bitfield's base word is assembled from byte reads that went unchecked, so a truncated frame read at an application-chosen offset answered with the bytes that followed it and set overwrote them, leaving half a word behind when it raised; and a byte_array, byte_array_where or byte_slice whose ~size is known only at decode time wrote the value's own length, running over the fields after it with no error. Both spans are now checked first (#262, #288, @samoht)
• Codec.set on a sub-codec, casetype or array field no longer leaves a refused value in the buffer. The writer laid the value down and only then checked it, against the documented promise to leave the buffer untouched; those writers now restore the bytes they replaced. Codec.encode is not transactional and now says so: after a failed encode the destination holds a partial record (#280, @samoht)
• Codec.get and Codec.set handle an optional or optional_or field whose gate is fixed at construction, a shape Codec.decode reads without trouble and get used to raise Failure "build_field_reader: unsupported type" on. A runtime gate stays unsupported and refuses with Invalid_argument naming the shape (#280, @samoht)
• Codec.size_of_value no longer answers for a write that cannot happen or under-reports a field's width: a casetype value no case projects raises Invalid_argument as encode does, and Codec.set refuses before writing, while a fixed-size byte_array, byte_array_where or byte_slice reports its declared width, so a buffer allocated from the answer exactly fits the bytes written (#238, #296, @samoht)
• Unexpected_eof reports byte counts, not buffer positions, and counts only bytes the parse was handed. The expected and got fields carried absolute positions at several sites, so the pair shifted with the offset the frame was read at and could come out equal to each other; and a value inside a nested region or a repeat budget was sized from whatever the buffer held past the region, so the reported shortfall named bytes the parse never claimed (#285, #326, #332, @samoht)
• A parse error names the field's own offset, so a caller locating the field from at no longer reads or rewrites the wrong bytes. A failing field ~constraint_, an enum membership failure and a staged read through a bitfield reported the enclosing record's base, and a lookup tag out of range reported byte 0 whatever offset the frame was read at. A dynamic layout still reports the record base (#306, @samoht)
• A failure under a size or offset expression reports its own error. A validation error or an exception from a map callback came out as an end of input on a buffer that was long enough, and a byte span whose size expression went negative escaped as Invalid_argument "String.sub"; truncation is now raised by the reads themselves, naming the missing span (#267, @samoht)
• Field.repeat, array, array_seq and nested hold to their declared size on decode and encode. A repeat consumed or emitted more or less than its byte budget, the array encoders accepted a short value and left zero-filled phantom elements, and nested let its inner value consume less than the region. Use nested_at_most where trailing region padding is intentional; it remains permissive and zero-pads on encode (#238, @samoht)
• Codec.v rejects duplicate field names, and distinct parameter handles with the same name. Name-based field access and parameter environments silently aliased the first declaration (#238, @samoht)
• A constant size or length expression such as Expr.(int 2 + int 2) behaves exactly like the literal int 4. It used to slip past every check and fast path that looks for a literal size, so such a field encoded without its exact-length check, reported its codec as variable-size, and bypassed the uint 1-7 size guard (#259, @samoht)
• Operations on the same parametric codec are safe to re-enter and interleave across fibers with different Param.env values: embedded codecs, casetype bodies, fixed-size wrappers and repeated elements no longer read another operation's field sizes. An env built for a different codec is rejected first, on encode, decode, validation and staged getters alike, including where the parameter counts happen to match (#237, #239, @samoht)
• A field wider than the target's native int no longer decodes or encodes wrong under wasm_of_ocaml and js_of_ocaml. A bitfield touching bit 31 of its base word dropped that bit, a 4-byte signed field dropped the top bits on the way in and sent the frame back out as different bytes with the encode-side range check disabled, and a NaN passed a float64 is_finite check. The emitted 3D refinements are unchanged (#232, #321, @samoht)
• A value too wide for the native int reports a typed Value_out_of_range, or a contextual Invalid_argument naming the case index or parameter, where reading a uint32, uint63 or uint field, projecting a casetype and Param.bind used to leak a bare Optint Failure. The validation slots mirror only values the int can hold, as uint64 already did (#232, #237, @samoht)
• uint32 and uint64 values order as unsigned numbers on every target. Their carriers compared the top bit as a sign, so the largest value of each ranked below 1 (#322, #323, @samoht)
• A shape with no valid encoding or no EverParse projection is refused at construction with a clear error, rather than building and misbehaving later: a Field.repeat element whose greedy tail sits inside a sub-codec and so swallows every element after it, a zero-size byte span as an array or Field.repeat element, an optional field that can expose a consume-rest payload with another field after it, and a field * constant byte size whose field <= bound constraint lets the maximum reach EverParse's 2^32 limit. Products without that conclusive shape stay deferred to EverParse, avoiding speculative rejections (#237, #239, #256, #266, #280, @samoht)
• A Field.repeat element that turns out to consume no bytes reports an end-of-input error from Wire.of_string and Wire.of_bytes instead of looping forever (#256, @samoht)
• Wire.Ascii diagrams render every expression a size or constraint can contain. Bitmasks, shifts, division, casts, conditionals, sizeof and parameter references used to print as a bare ?, so a trailing payload sized by Wire.rest_bytes showed up as (? - ? bytes) (#257, @samoht)
• Generated C no longer enforces the wrong specification when names collide. Two codecs whose artifacts differ only in the leading capital or in an uppercase run, and two codecs of a family that declare different types under the same name, are refused with a clear error instead of the second write winning and leaving the shadowed codec's stubs validating against another codec's spec; and two domains building descriptions at once can no longer mint the same synthesised byte_array_where element name, whose 3D typedef was then emitted twice. Types declared identically by several codecs still collapse to a single declaration (#255, #266, #272, @samoht)
• The generated FFI stubs no longer lose field values. The parse callback collected the boxed arguments in a bare value array, which is not a GC root, so a minor collection triggered by a later box could move or reclaim an earlier one and the continuation silently received garbage. CAMLlocalN now roots them (#272, @samoht)
• The emitted EverParseEndianness.h compiles on a target that is neither Linux/glibc nor Apple. The __BYTE_ORDER__ branch was spliced only when the header was absent, which it never is, so the header a cross-compiled standalone archive gets stopped at #error "Unsupported platform" (#272, @samoht)
• EverParse generation no longer interprets output paths, executable paths and schema filenames that contain spaces or shell metacharacters as commands (#239, @samoht)
• A package's committed .3d specs, dune.inc and EverParse C can no longer go stale. dune runtest diffs the specs and rules against freshly generated ones and rechecks an installed <Name>.provenance stamp holding the stdlib BLAKE2b-256 digest of the .3d the C was built from, no EverParse needed. The committed schema is a plain source rather than a promoted build target, so drift fails runtest with a promotable report and dune promote writes the new bytes after an intended codec change (#235, #317, @samoht)

Removed

• Breaking: Remove uint63 and uint63be. An 8-byte field spans more values than their Optint.Int63.t carrier holds, and decoding masked the difference away instead of failing, on both the OCaml and the generated C path. Use uint64 / uint64be instead, whose carrier is exactly the eight bytes on the wire, and convert an existing value with Optint.Int63.to_int64. Wire.uint ~size is unaffected: its carrier is exact for the one to seven bytes it accepts (#316, @samoht)
• Delete the dependency on eio. Nothing in the library used it, it was only ever linked, unused, into wire.diff (#233, @samoht)

1.1.0

Added

• Wire.Expr.land64 is an int64 bitwise AND for masking a full-width uint64 field inside a constraint, and ~self_int64 / Field.int64 now accept a map-decoded uint64 field. A sign-magnitude offset bound such as a bsdiff seek (land64 self mask <= max) is now expressible and projects to a 3D refinement EverParse verifies (#227, @samoht)

Changed

• The standalone c/ archive (3d mode) now builds and installs in every dune context through that context's own toolchain, so a cross build produces and installs a target-native verified parser instead of skipping the C. The build uses %{ocaml-config:c_compiler}, the ocaml-config partial linker, and the binutils that compiler resolves (-print-prog-name), so a cross cc finds the target's objcopy/ar. EverParseEndianness.h gains a freestanding __BYTE_ORDER__/__builtin_bswap* branch so it compiles for an OS-less target (a unikernel defines neither __linux__ nor __APPLE__). Only C regeneration and the agree differential test stay host-gated (#231, @samoht)
• uint32/uint32be now decode to Optint.t and uint63/uint63be to Optint.Int63.t rather than a native int, so a value with bit 31 (or the high half) set is no longer silently truncated on a target whose int is narrower than 63 bits (js_of_ocaml, wasm_of_ocaml); a TCP sequence number now round-trips there. Read such a field with Optint.to_int / Optint.to_int32; a uint32 used as a size parameter must become int32be (#226, @samoht)
• Decode errors are redesigned. parse_error is now a { at; field; kind } record instead of a flat variant: at is the failing field's byte offset, field the root-to-leaf path of field names to it, and kind a closed error_kind (Unexpected_eof, Invalid_enum, Invalid_tag, Missing_terminator, Non_zero_padding, Value_out_of_range, Constraint_failed), so a failure deep in a nested struct is both locatable and matchable. Constraint_failed of string becomes Constraint_failed { which; value }, naming the predicate that failed and carrying the offending field's value. Validation_error is removed (Codec.validate and the _exn decoders raise the single Parse_error), and the type gains equal_parse_error / compare_parse_error, pp_error_kind, and public parse_error / eof constructors. Match on e.kind, read e.at, replace Validation_error with Parse_error, and wrap Wire.parse_error in your own variant for domain errors (#219, @samoht)
• Decoding a record codec with 17 to 32 fields no longer allocates a short-lived closure on each decode: the closure-free decode ceiling that 1.0.0 raised to 16 fields now extends to 32. Real protocol headers cross the old limit routinely, such as an 18-field TCP header with its flag bits broken out, or telemetry and IPv4 headers in the high twenties, which now allocate only the record they return. A codec wider than 32 fields keeps the recursive fallback (#214, @samoht)

Fixed

• A standalone Wire_3d archive no longer exposes its raw <Base>Validate* validators to C callers. Those entrypoints skip a StartPosition <= InputLength bound check, so a direct caller passing StartPosition > InputLength underflowed the read span and went out of bounds. The install now ships only the checked <Base>Wrapper.h header (not <Base>.h), and the archive build localizes every symbol except the <Base>Check<Codec> wrappers, so the raw validators are neither declared nor linkable; the wrapper validates from position 0 and is the safe public API (#228, #229, @samoht)
• Codec values are now safe to share across domains: their validation scratch and parameter backing are per-domain, so encoding, decoding, or validating one codec concurrently no longer corrupts the result (#223, #224, @samoht)
• A generated FFI parser (parse buf off) now raises Invalid_argument when off is negative or past the end of buf, instead of reading out of bounds. The C stub computed its length as Bytes.length buf - off in unsigned arithmetic, so an out-of-range off (which can carry a length or offset field parsed from untrusted input) underflowed the length into a huge span and pushed the read pointer past the buffer before the validator ran. An in-range off is unaffected (#222, @samoht)
• A non-zero byte in an all_zeros padding field now reports the same error whether the field is decoded through a struct (Codec.decode / Codec.validate) or directly (Wire.of_string). The struct path used to report a stringly Constraint_failed that dropped the byte offset while the direct path reported the typed All_zeros_failed with the offset; both now report All_zeros_failed carrying the offset (#220, @samoht)
• A casetype whose discriminant matches no case now fails with the typed Invalid_tag carrying the tag value, the same class as an out-of-range lookup index, instead of a stringly Constraint_failed (#220, @samoht)
• A non-integer field referenced where an integer is required (a schema mistake, such as a float used as a length) now raises Invalid_argument rather than a parse error, keeping schema errors distinct from malformed input (#220, @samoht)

1.0.0

Added

• Field.action returns a field's action (the ?action passed to Field.v), completing the field accessor set alongside Field.constraint_ and Field.doc. Action.pp and Param.pp pretty-print an action block and a parameter, matching the existing Field.pp (#189, @samoht)
• Wire.enum_open name cases base is an open enumeration: it names the known values for documentation but accepts any value. Unlike Wire.enum / Wire.variants, it does not reject an unlisted value (no decode Invalid_enum, and the field projects as its base scalar with no membership refinement), which is what an open value set (a protocol field that may carry unknown or future codes) needs. The known codes are still emitted as a 3D enum declaration, so they stay documented in the generated .3d (#166, @samoht)
• The doc pipeline's differential agree.c is now derived from the codecs alone: it computes each validator's name and parameter types from the Wire definitions instead of reading the EverParse-generated <Name>Wrapper.h. The self-check therefore regenerates as pure OCaml, with no EverParse needed and no chance of drifting from the codec, and its build rule is split out from the committed C. Wire.Everparse.Raw.input_param_c_types exposes the per-parameter C types this relies on (#167, @samoht)
• Wire.Expr.if_then_else cond t e builds a conditional value expression (the 3D ? : ternary), so a size or constraint can depend on another field, e.g. a 16-bit length where 0 means 65536: if_then_else Expr.(len = int 0) (int 65536) len. The underlying constructor was previously reachable only through the wrapped internal module (#164, @samoht)
• The doc pipeline's differential harness now covers parameterized codecs: the corpus oracle binds each codec's Param.input values and the generated agree.c passes the same values to the EverParse validator, so a length-bound or otherwise parameter-dependent frame (e.g. a CCSDS TC / AOS / TM / USLP layout) is checked end to end rather than skipped. Wire.Param gains bind_by_name, which binds an input parameter by name without its typed handle (#163, @samoht)
• Wire.Field.v takes an optional ?self_int64, and Wire.Field.int64 / Wire.Expr.int64 build full-width 64-bit field constraints. This lets schemas constrain domains such as signed-magnitude uint64 values without truncating the field through OCaml's native int (#160, @samoht)
• Wire.Field.v takes an optional ?doc (read back with Wire.Field.doc): a free-text note, such as an RFC section, that the documentation projection renders as a /* ... */ comment above the field in the generated 3D. A protocol spec can now cite the standard each individual field comes from, not just the struct as a whole, and EverParse accepts the comment (#157, @samoht)
• The standalone projection pipeline (Wire_3d.main ~mode:Standalone) now auto-generates a differential self-check: dune runtestfuzzes inputs, records whether the OCaml codec accepts each, and replays them through the EverParse-generated C validator, failing on any input the two decide differently. This catches a doc projection that drifts from the codec (a wrong bit order, a constraint that means something else over the wire type), which nothing checked before since the doc validator carries no FFI. The build also produces an installedlib.aarchive of the validator. NewWire_3d.generate_corpusandWire_3d.generate_agree` expose the two halves (#158, @samoht)
• Wire.Codec.v takes an optional ?doc (read back with Wire.Codec.doc): a free-text note, such as an RFC citation, that the documentation projection renders as a /*++ ... --*/ comment on the codec's 3D typedef. The generated spec then documents which standard each protocol struct comes from, and EverParse accepts the comment (#155, @samoht)
• Wire_3d's documentation helpers (generate_doc, generate_dune_doc, and main ~mode:Standalone) take an optional ?namethat sets the generated.3d/.cfile base independently of the opam~package, so a package like ocaml-tcpcan emit aTcp.3d` spec while still installing under its own name (#154, @samoht)
• Wire.Everparse.project ~mode:StandaloneandWire.Everparse.writeproject a codec, or a whole family of codecs, to a clean.3dwith no FFI scaffolding: enums render as named 3D enum types, types shared across codecs are emitted once, and a protocol family lands in one readable.3d`. The result doubles as a protocol specification and as input to EverParse, which compiles it to a standalone verified C parser with no FFI (#151, @samoht)
• Wire.Codec.rename returns a codec with a new name, leaving its wire encoding and field constraints unchanged, so a generically built codec can be given a unique, meaningful name before code generation (#121, @samoht)
• Wire.nested / Wire.nested_at_most now accept a composite inner (a Wire.array, or another nested region), and a Wire.casetype field's case body may be such a region; both round-trip and generate a verified EverParse validator (#109, @samoht)
• An embedded sub-codec (Wire.codec c used as a field or Field.repeat / array element) that takes Param.input parameters now works: the outer codec exposes the sub-codec's input params as its own, so Codec.env / Param.bind reach them and the values are threaded into the sub-codec on encode, decode, and projection (#108, @samoht)
• A Wire.casetype used as a Field.repeat element may now have a bitfield case body, alongside the scalar, byte-span, NUL-terminated, and sub-codec bodies already allowed (#105, @samoht)
• Field.optional and Field.optional_or now accept a variable-size inner, so an optional field can be a length-prefixed string or a whole sub-message, not just a fixed-width value, and generate a verified EverParse validator (#88, #133, @samoht)
• Wire.zeroterm and Wire.zeroterm_at_most ~size for NUL-terminated strings: the bytes up to a terminator, optionally bounded to a fixed-size region (#77, @samoht)
• Wire.casetype now accepts any tag type ('k typ, not just int), so a string-discriminated union (a byte_array-tagged casetype, as in many SSH messages) is expressible (#49, @samoht)
• A Wire.casetype case body can be an embedded sub-codec, including one that ends in all_bytes (#50, @samoht)
• Field.repeat now supports variable-size elements, e.g. a list of length-prefixed sub-messages (#51, @samoht)
• Wire.Codec.size_of_value: the encoded byte length of a value (#58, @samoht)
• Wire.casetype and Wire.nested ~size can be used as Codec fields (#47, @samoht)
• Add Field.optional / Field.optional_or / Field.repeat / Field.repeat_seq (#46, @samoht)
• Add Wire.rest_bytes for trailing "rest of buffer" fields, plus direct all_bytes / all_zeros support as Codec fields (#44, @samoht)
• Add signed integers int8 / int16(be) / int32(be) / int64(be) (#42, @samoht)
• Add IEEE 754 floats float32(be) / float64(be) and Wire.is_finite / Wire.is_nan predicates (#42, @samoht)
• Add Wire.byte_array_where ~size ~per_byte for byte spans with a per-byte refinement (#40, @samoht)
• Add Codec.validator_of_struct / validate_struct / struct_size_of / struct_min_size (#37, @samoht)
• Add Codec.slice_offset / Codec.slice_length (#37, @samoht)
• Add Wire.codec type alias for 'r Codec.t and Wire.pp_value (#39, @samoht)

Changed

• Encoding a variable-length byte field (byte_slice, byte_array, zeroterm, all_bytes, ...) no longer allocates: on a flambda-off switch the writer rebuilt two short-lived closures per field on every encode. Encoding a record made of scalars and var-bytes fields is now allocation-free (#213, @samoht)
• Decoding a record codec with more than 8 fields no longer allocates a short-lived closure on each decode: the constructor is now applied in one saturated call for codecs of up to 16 fields. On a flambda-off switch this partial application was the dominant cost of decoding a wide header (CLCW at 13 fields, CFDP at 14), which now allocates only the record it returns (#212, @samoht)
• The EverParse projection API is consolidated to two entry points. The two projections, previously Wire.Everparse.schema and Wire.Everparse.doc, are now Wire.Everparse.project ?mode, where `Standalone (the default) emits a clean .3d that EverParse compiles to a standalone verified C parser (the production output, which also reads as a spec) and `Ffi emits the OCaml-callable bridge with WireCtx extern callbacks. Writing is likewise one Wire.Everparse.write ?mode (replacing write_3d and write_doc). The struct-level entry points struct_of_codec and schema_of_struct move under Wire.Everparse.Raw (as struct_of_codec and project_struct), and Wire_3d.main's mode is now `Standalone rather than `Doc (#210, @samoht)
• A long ?doc note on a field or codec (an RFC citation, say) now wraps across several comment lines in the generated .3d instead of rendering as one line past 80 columns, so the generated spec stays readable (#191, @samoht)
• A package generated by the Wire_3d pipeline now treats its EverParse C (<Name>.c, the wrappers, EverParse.h) as committed source. A plain dune build or dune test uses the committed C and never invokes 3d.exe, and fails loudly if the C was never committed; set BUILD_EVERPARSE=1 to regenerate it (the rule then promotes the result back into the tree). Previously the rule regenerated whenever the C was missing, which silently required EverParse on an ordinary build and let committed C drift unnoticed. The pure-OCaml .3d and agree.c are unaffected and still regenerate on demand (#168, @samoht)
• Wire_3d.main now takes packed codecs (Wire_3d.pack codec) and a mandatory ~mode:[Ffi | Standalone], so every gen.ml states what it emits. `Ffi keeps the per-codec FFI parsers; `Standalone emits one FFI-free <Package>.3d specification and a single standalone <Package>.c parser for the whole package, through the new Wire_3d.generate_doc and Wire_3d.generate_dune_doc. Migrate a gen.ml by replacing [schema c; ...] with ~mode:Ffi [pack c; ...], or ~mode:Standalone for the single-file output (#152, @samoht)
• Reading or writing a uint32 or uint63 field now stays in the native int instead of round-tripping through a boxed Int32 or Int64. The boxing surfaced as per-field allocation in tight decode and encode loops; field access is now allocation-free regardless of how the compiler optimises the surrounding code. Pure speedup, no API change (#150, @samoht)
• Wire.Codec.decode no longer allocates a fresh validation buffer on every call: each codec reuses a single buffer across decodes, so decoding the same codec in a loop allocates a constant amount instead of growing with the number of decodes. Pure speedup, no API change (#149, @samoht)
• Decoding a Field.repeat over a Wire.casetype (the DHCP / TCP option-list shape) no longer allocates a closure and a boxed length per element, so decode allocation no longer grows with the number of elements. Pure speedup, no API change (#148, @samoht)
• The build rules wire.3d generates (the dune.inc and the wrapper in its setup example) now use a 3d alias instead of the generic gen, so dune build @3d regenerates the .3d files and EverParse C parsers. Update any dune build @gen invocation accordingly (#146, @samoht)
• Wire.of_reader now rewinds on failure: every byte consumed by a failed decode is pushed back, restoring the reader to its position before the call, so the caller can retry with another description or after more input arrives (#145, @samoht)
• Wire.of_reader now consumes only the bytes of the decoded value and leaves the rest on the reader, so several values can be decoded back-to-back from the same reader. Previously the first call drained the whole reader. Types that extend to the end of input (all_bytes, all_zeros) still consume the whole stream (#144, @samoht)
• The Wire.Everparse.plug_field record fields lose their pf_ prefix (pf_name is now name, pf_idx is idx, and so on). Update custom plug generators accordingly (#144, @samoht)
• Codecs that share a synthesised type (an enum, or a refined-byte or element-wrapper struct) can now be linked into one binary, so full protocol stacks built from per-codec parsers (Ethernet, IPv4, TCP, ...) link cleanly. The Wire_stubs-generated FFI now compiles each codec's C as its own unit and links them; use the new Wire_stubs.build_codec_archive to build the link archive (#135, @samoht)
• Field.repeat and Wire.array over a Wire.casetype now raise Invalid_argument at construction when a case body has no per-element projection (a nested region, array, or optional), instead of building a codec that fails later at decode (#105, @samoht)
• Wire.default (a casetype's default branch) no longer takes a fixed ~tag; instead it threads the matched discriminator through inject and project, so an arbitrary unclaimed tag round-trips. inject is now 'k -> 'w -> 'a (it receives the matched tag along with the body) and project is now 'a -> ('k * 'w) option (it returns the tag to write back), so a default branch can recover and re-emit the tag it caught (e.g. a DHCP or TCP option code). Migrate by taking the tag in inject (fun _tag body -> ... to ignore it) and pairing it in project (#100, @samoht)
• Decoding a struct with variable-size fields (byte_slice, byte_array, or a repeat sized by a cross-field expression) no longer allocates on each field access. Pure speedup, no API change (#81, @samoht)
• Remove Wire.optional / Wire.optional_or / Wire.repeat / Wire.repeat_seq from the typ-level surface; use the matching Field.* combinators instead (#46, @samoht)
• Rename Wire.decode_* / Wire.encode_* to of_string / of_bytes / of_reader / to_string / to_bytes / to_writer; add _exn twins that raise on parse error (#39, @samoht)
• Fold Codec.decode_with into Codec.decode via ?env; split into Codec.decode (result) and Codec.decode_exn (raises) (#39, @samoht)
• Speed up Codec.bitfield ~5% (#37, @samoht)
• Drop the pe_ prefix on Types.param_env fields: they are now codec_id, slots, bound. Callers using Param.bind / Param.get are unaffected; code that pattern-matches the record or reads fields directly needs to update (#63, @samoht)
• Codec.raw_encode now returns the offset after the written bytes instead of unit. Codec.encode is unaffected (#65, @samoht)

Documentation

• Add odoc cross-reference links and the doc comments that were missing across the public interfaces, and expose a pp printer on UInt32, UInt63, Param, and Wire.Diff (#73, @samoht)
• Type-check README.md and every public .mli under mdx (#39, @samoht)

Fixed

• The generated C <Name>Check<Codec> wrapper now validates the whole buffer: it returns FALSE unless the validator consumed every input byte, so a valid record followed by trailing bytes is rejected instead of accepted as a valid prefix. The differential corpus oracle applies the same whole-buffer rule, so the OCaml and C verdicts still agree. The raw <Name>Validate<Codec> entry point keeps its prefix semantics and returns the consumed position. wire now depends on re at runtime, previously a test-only dependency (#215, @samoht)
• Wire.Everparse.project now rejects a codec that cannot project to 3D when the schema is built, not later when it is rendered. A constraint with no 3D projection (a field_pos, or a subtraction or multiplication over a field) used to build a schema that only raised once passed to to_3d, so the projection was not a reliable projectability check (#209, @samoht)
• Codec.validate now enforces every check Codec.decode does, for any codec. It used to skip a field's own decode-side checks (enum and variant membership, a lookup index bound, a refined or NUL-terminated span, an embedded codec or array element's constraints) whenever the struct had no top-level constraint or where, so validating untrusted input accepted values decode rejects and a following zero-copy Codec.get trusted them. A codec with nothing to check (a header of plain scalars and byte spans) still validates without allocating, so validating before a batch of get calls on a hot path stays cheap (#207, #211, @samoht)
• A field constraint that subtracts or multiplies a field (such as Expr.(a * b = int 0) or Expr.(a - b >= int 5)) is now rejected at projection. Such an expression under- or overflows the field's narrow width, which EverParse refuses to verify, leaving the codec with no validator; unlike addition (widened to avoid overflow), neither has a sound projection. A constant Sub / Mul and additive field arithmetic are unaffected (#206, @samoht)
• Codec.validate now enforces an all_zeros padding field, rejecting a non-zero byte exactly as Codec.decode does. The zero check lived only in the decode reader, so validating a frame with tampered padding succeeded and a following zero-copy read accepted it (#205, @samoht)
• Codec.validate now runs decode's structural bounds check for every codec, including one with no field constraints or where. It used to be a no-op for a constraint-free codec, so validating a truncated buffer succeeded and a following zero-copy Codec.get read out of bounds; the documented safety gate now rejects a short buffer (#204, @samoht)
• A signed field equality constraint (an int8 field whose ~self_constraint is Expr.(s = int k)) now projects to the two's-complement byte the unsigned validator reads, and folds to a constant when the target is outside the signed range. The raw constant was compared against the unsigned projection, so the generated C and Codec.decode disagreed on the same byte (#203, @samoht)
• A byte_slice whose resolved size is negative, for example a Sub on a length field that underflows on crafted input, now fails with a Codec parse error instead of escaping Codec.decode with a raw Invalid_argument. The slice read skipped the bounds check the other byte spans run and crashed (#202, @samoht)
• A Wire.casetype whose tag is a uint ~size value, or an enum over a big-endian base, is now rejected at construction. Neither projects to a 3D type the dispatch can name, so the codec built without a verified validator. A fixed-width integer, bitfield, or little-endian / 1-byte enum tag is unaffected (#201, @samoht)
• A non-trivial Wire.where used as a Wire.casetype case body is now rejected at construction. Such a refinement projects to case k: T { cond } v;, which is not valid 3D (a case body takes no refinement), unlike a top-level field where, so the codec had no verified validator (#200, @samoht)
• Wire.bits ~width is now validated against its base word: a width above the base size (e.g. bits ~width:9 U8) or below 1 is rejected at construction. Such a field had no faithful wire meaning, and the OCaml shift and the EverParse-generated validator read different values from the same bytes (#199, @samoht)
• Field.repeat over a zero-width element (Wire.empty) is now rejected at construction, like Wire.array already was. A byte-budget list of a 0-width element does not extract through EverParse, so the codec had no verified C validator; the error now fires when the schema is built rather than producing an unverifiable spec (#198, @samoht)
• A closed enum (or variants) used as an array element, a repeat element, or an optional inner now enforces its value set in the EverParse-generated C validator, matching Codec.decode. Previously only a byte-wide enum array element carried the membership check; a wider or big-endian array element, a repeat element of any width, and an optional inner accepted codes outside the named set that the OCaml decoder rejects with Invalid_enum, so the two disagreed on the same input (#197, @samoht)
• The documentation pipeline's differential agree.c now predicts each EverParse validator-wrapper symbol with Wire_3d.pascal_case, a transcription of EverParse's own pascal_case mangling, computed as pascal_case (module ^ "_check_" ^ codec). The old code glued everparse_name-normalized parts, which kept a capital after a digit (TPM2B gave TpmCheckTpm2B where EverParse emits TpmCheckTpm2b), so agree.c referenced a symbol that did not exist and failed to compile under -Werror. Regenerate a package's agree.c to pick this up (#196, @samoht)
• The dune.inc rules generated by the Wire_3d pipeline now invoke the generator through %{exe:gen.exe} and use plain dune actions instead of shell (system ...) commands. The old rules ran ./gen.exe, which relied on dune placing a gen.exe symlink in the action's working directory; under sandboxing that is not guaranteed, so the build could fail with gen.exe: not found. The corpus, the differential agree driver, and the validator archive are now built as ordinary targets and run directly. Regenerate a package's dune.inc (dune build @<pkg>/c/3d) to pick this up (#195, @samoht)
• A codec whose name has a lower-case segment after an underscore (such as Grpc_message) now generates C identifiers that match the ones EverParse emits. Wire_3d.everparse_name capitalized only the first segment, yielding Grpcmessage where EverParse produces GrpcMessage, so the generated FFI stubs and the documentation differential harness referenced a name that did not exist and failed to link. Every underscore-separated segment is now capitalized (#193, @samoht)
• A field constraint that adds narrow unsigned fields (such as a + b <= 10 over two uint8 fields) now projects to a .3d EverParse verifies. The sum was emitted at the field's own width, which EverParse refuses to verify because it can overflow, while Wire.Codec.decode computes the same sum in OCaml's wide native int. The addition's operands are now widened to 64-bit so the generated C validator and the decoder agree (#188, @samoht)
• An array of an open enum (Wire.enum_open) now validates identically in the OCaml decoder and the EverParse-generated C validator: both accept any element value. The C validator used to reject element values outside the named codes (it constrained each element to the enum's named set), while Wire.Codec.decode accepted them, so the two disagreed on which buffers are valid (#187, @samoht)
• An enum / variants over a big-endian base (e.g. enum ... uint16be) now projects to a .3d EverParse accepts. It was emitted as a UINT16BE enum declaration, which EverParse rejects (it types the integer constants as the native width: "Expected UINT16BE, got UINT16"). A big-endian-based enum now projects as its base scalar with a membership refinement (closed) or bare base (open), with no enum declaration (#185, @samoht)
• Wire.array / Wire.array_seq now reject a zero-width element (empty / unit) at construction. Such an array carries no bytes and projected to a zero-size 3D array EverParse rejects; it is a degenerate shape and is refused up front (#184, @samoht)
• Wire.Codec.v now rejects, at construction, a non-last field that is a Wire.casetype with a case body ending in a greedy field (all_bytes / all_zeros). If that case is selected the greedy tail consumes the rest of the buffer, starving the following field, so the record failed to round-trip while construction silently accepted it. The greedy-must-be-last check now looks through casetype case bodies, as it already does through an embedded sub-codec (#183, @samoht)
• Wire.Codec.validate on a buffer too short to hold the fields a check reads now fails cleanly instead of raising Invalid_argument. A where or field constraint may read a field whose offset depends on a length read from the buffer; decode bounds-checks the buffer first, but validate ran the check kernel directly, so a short buffer read out of bounds and crashed (#181, @samoht)
• A zero-length Wire.byte_slice now decodes to an empty slice instead of raising Invalid_argument. The slice constructor rejects a zero length, so a byte_slice whose size resolved to 0 crashed the decoder rather than yielding the empty slice (#180, @samoht)
• Wire.Codec.v now rejects, at construction, a non-last field whose type is an embedded sub-codec ending in a greedy field (all_bytes / all_zeros). Such a tail consumes the rest of the buffer with no boundary, so it silently swallowed the following field's bytes and the record failed to decode. The greedy-must-be- last check already rejected a bare greedy field that is not last; it now looks through an embedded sub-codec too (#179, @samoht)
• Wire.of_string (and the other typ-level entry points) now return a clean Error on a truncated input to a variable-size codec, instead of raising Invalid_argument. Computing the codec's span reads its length and gate fields up front; on a buffer too short to hold them, that read ran off the end and escaped as an out-of-bounds exception rather than a parse error. Codec.decode already guarded this; the typ-level path now does too (#178, @samoht)
• Wire.of_string (and the other typ-level entry points) now accept an unlisted code in a Wire.enum_open field, matching Wire.Codec.decode. The typ-level decoder kept the closed-enum membership check regardless of the enum_open flag, so it raised Invalid_enum on a value the codec accepts, an open enum that behaved as closed on that path (#177, @samoht)
• Wire.Codec.decode now rejects a parametric codec whose env is missing or leaves an input param unbound, raising Invalid_argument (naming the param) the way Codec.encode already does. Decoding without binding a parameter used to resolve a parameter-driven field size to 0 and silently truncate the field; the binding precondition is now enforced up front on both sides (#176, @samoht)
• Wire.Param.bind_by_name now drives a parameter-dependent field size on decode, not only where clauses and constraints. A field whose size comes from a parameter (a byte_array, byte_slice, or uint_var sized by Param.expr) read as zero bytes when its parameter was bound by name, silently truncating the field and misaligning everything after it; only the typed Param.bind worked. Both binders now resolve parametric sizes identically (#175, @samoht)
• A signed integer field's ordering constraint (e.g. int8 x with x < 100) now projects soundly. A signed field becomes an unsigned UINT* in 3D, so the refinement was emitted as an unsigned comparison and the verified C validator disagreed with the OCaml decoder on bytes whose sign bit is set (byte 200 is the signed value -56: accepted by OCaml, rejected by C). The ordering is now rewritten to its two's-complement unsigned form. A float field ordering constraint, which has no faithful unsigned projection (IEEE bit patterns do not order as unsigned), is rejected when the codec is projected (#174, @samoht)
• A Wire.where placed on a container element (an array or repeat element, or an optional inner) is now rejected at codec construction with a clear error. EverParse cannot express a refinement on an array or optional element that references an outer field, so such a where produced a codec whose generated .3d did not compile while OCaml decode silently ignored the constraint. A where is supported as a top-level field refinement; move the constraint onto the field itself or a codec ~where (#173, @samoht)
• The doc pipeline's differential self-check (agree.c) now links for a codec whose name has interior consecutive capitals. EverParse normalizes such a name in the validator symbol (SpaceOSFrame -> SpaceOsframe), but the harness built the symbol from the raw name, so the generated check called an undeclared function. The name now goes through the same normalization, which also collapses a consecutive-capital run anywhere in a name, not only at the start (#171, @samoht)
• The doc pipeline's differential self-check (agree.c) no longer false-reports a mismatch for a codec with a large payload. The generated reader held each corpus line in a buffer one char short of two hex digits per input byte, so an 8 KB payload (16384 hex chars) truncated the line and misparsed the verdict. The buffer is now sized to the input width (#172, @samoht)
• Wire.Codec.decode and Wire.Codec.validate now enforce a constraint written as Wire.where cond t on a field, and any field ~action. Such a where was projected into the generated .3d (so the EverParse C validator rejected violating input) but was silently dropped on the OCaml side, so OCaml accepted what the verified C rejects; and Codec.validate skipped field actions that decode ran, so the two disagreed. Decode and validate now share a single validation path and enforce identical semantics (#169, @samoht)
• A Wire.casetype that switches on a Wire.enum tag now projects to a 3D schema EverParse accepts: each case label is emitted as the enum constant name (case InteriorIndex:) instead of the raw integer (case 2:), which EverParse rejected as not a member of the enumerated type (#162, @samoht)
• Wire.Expr.( = ) and Wire.Expr.( <> ) are explicitly re-exported from the expression language, so equality in a local Expr.(...) open builds Eq / Ne constraints rather than depending on the surrounding equality binding (#159, @samoht)
• A codec whose name contains a capital V (e.g. VeritySuperblock) now generates its C parser. EverParse names the validator <Name>Validate<Name>, and the name reader stopped at the first V, so C generation failed for any such name (#143, @samoht)
• A Wire.enum field now enforces its membership in the EverParse-generated C validator, rejecting values outside the named cases exactly as Codec.decode does (raising Invalid_enum), including for an enum nested inside a sub-codec or record. Previously the verified C accepted out-of-range values the OCaml decoder rejects (#131, @samoht)
• Decoding a Wire.enum through the Codec API now rejects a value that is not one of the named cases (raising Invalid_enum), on a scalar field and on every array or repeat element, matching the EverParse validator and the Wire.of_string path. The Codec decoder previously accepted any value (#130, @samoht)
• A Wire.lookup field now enforces its index bound in the EverParse-generated C validator, on a scalar field and on every Wire.array / Field.repeat element, so the validator rejects out-of-range indices exactly as the OCaml decoder does. Previously the verified C accepted indices the decoder rejects (#126, @samoht)
• A Wire.nested / nested_at_most field, and a Wire.byte_array_where span inside one, now generate a verified EverParse validator. Previously any codec with such a field failed schema generation (#99, #132, @samoht)
• Wire.rest_bytes now generates a verified EverParse validator. Previously any codec with a rest_bytes field failed schema generation regardless of width (#117, @samoht)
• A codec with a Wire.uint63 / uint63be field now generates a verified EverParse validator (projecting to the 8-byte UINT64). Previously schema generation failed and the codec had no verified C parser at all (#125, @samoht)
• A codec mixing signed-integer or float fields of different widths (e.g. a float32 then a float64, or an int8 then an int32) now generates a verified EverParse validator. Previously such a codec had no verified C parser (#127, @samoht)
• A Wire.enum used as a Wire.array / Field.repeat element (or inside an optional or sized region) now generates a verified EverParse validator. Previously such a codec had no verified C parser (#128, @samoht)
• A statically-absent Field.optional / Field.optional_or (~present:false) now generates a verified EverParse validator. Previously the codec had no validator at all (#137, @samoht)
• An Action.on_success ending in a conditional Action.return_bool (an Action.if_ with a return branch), and an Action.on_act whose body ends in Action.return_bool, now generate a verified EverParse validator (#139, @samoht)
• Projecting an expression (a ~where / field constraint / ~self_constraint) that uses a construct with no projectable form (a negative integer literal or field_pos) now raises a clear Invalid_argument instead of emitting C that EverParse rejects with a cryptic error. Every other operator (shifts, bitwise, casts, mod, div, comparisons, sizeof, sizeof_this) projects (#140, @samoht)
• The generated dune rule now compiles the EverParse C under strict C11 (-std=c11 -D_DEFAULT_SOURCE) instead of -std=c99, so the verified validators build on Linux glibc (the BSD endian helpers the C uses need _DEFAULT_SOURCE) (#134, @samoht)
• Decoding no longer raises Invalid_argument on adversarial input where a Field.repeat byte budget, or a variable field's cross-field size, exceeds the buffer; an oversized length now fails with a clean Parse_error (#117, @samoht)
• A cross-field length / offset / present expression that reads an integer beyond the native int range (a uint64 / int64 length over max_int), or reads a non-integer field, now fails the parse instead of silently reading 0. The old behaviour masked malformed input (#82, @samoht)
• A byte_array / byte_slice (or any field) whose ~size reads a Field.optional_or field no longer resolves that size to 0, decoding the span as empty (silent truncation) and raising a length mismatch on Codec.encode; it now reads the present-or-default value (#101, @samoht)
• A greedy field (all_bytes / all_zeros) reads the rest of the buffer, so it is now rejected with Invalid_argument anywhere it is not the final field: a non-last field of a codec, a Field.repeat / Wire.array element (or a sub-codec ending in one), or a Wire.casetype case body. It remains valid as the last field, the supported way to consume the rest (#107, #110, #111, @samoht)
• An embedded sub-codec's where clause and field constraints are now enforced when the codec is decoded as a field or element. They were silently dropped on the embedded path, so a value the sub-codec would reject standalone was accepted when embedded (#108, @samoht)
• Wire.array / array_seq now reject a non-fixed-width element (a Wire.nested region, a Wire.byte_array_where refined span, or a nested Wire.array) at construction with Invalid_argument. An element must be a scalar, a fixed byte span, or a fixed-size sub-codec, matching Field.repeat (#107, @samoht)
• Wire.array / array_seq / Field.repeat / repeat_seq over a sub-codec built only from byte-span fields (byte_array, byte_slice, a varint) now raise Invalid_argument at construction. A sub-codec with at least one fixed-size field is accepted as before (#115, @samoht)
• Wire.array / array_seq over a float, a signed integer, a uint63, or a Wire.where / Wire.map wrapping a fixed byte span no longer raise during 3D projection; all fixed-width scalars and wrapped byte spans now project (#116, @samoht)
• Field.repeat / repeat_seq now reject an element type that has no clean per-element projection (a sub-byte bits field, a refined or at-most byte span, all_zeros, or a nested array / nested) at construction with a clear Invalid_argument. Supported elements are unchanged: fixed-width scalars and byte spans, zeroterm, sub-codecs, and casetypes (#97, @samoht)
• A bitfield (Wire.bits / Wire.bit) is now rejected with Invalid_argument at construction as an element of Field.repeat / Wire.array / Wire.nested or as a Field.optional inner: a bitfield only exists packed inside a record, with no standalone wire form (#90, #98, #107, @samoht)
• A Wire.casetype whose case body is a NUL-terminated string (zeroterm or zeroterm_at_most) now encodes, decodes, and sizes correctly as a Field.repeat element, so a list of such tag-dispatched options round-trips (#103, @samoht)
• Field.repeat over a Wire.casetype element now encodes and decodes instead of raising. This covers DHCP-style options whose cases mix bare single-byte tags with length-prefixed bodies (#75, @samoht)
• Field.repeat / Wire.array over a fixed byte_array / byte_slice element (a list of n-byte chunks, e.g. fixed-size addresses) now encode, decode, and generate a verified EverParse validator; decoding previously raised Failure (#89, #92, @samoht)
• Field.repeat over a zeroterm element (a list of NUL-terminated strings within a byte budget) now encodes, decodes, and generates a verified EverParse validator; it previously raised Failure when decoding (#93, @samoht)
• Wire.array over a fixed-size sub-record (a Wire.codec element, e.g. an array of { x; y } points) now decodes instead of raising Failure (#96, @samoht)
• A Field.optional_or with a dynamic gate now generates an EverParse C validator that accepts the bytes Codec.encode produces; the two previously disagreed on the field's layout when the gate was false (#88, @samoht)
• A codec that embeds a variable-size sub-codec (Wire.codec, e.g. a length-prefixed string) as a field is now accepted by EverParse; it previously failed schema generation (#87, @samoht)
• A variable-size sub-codec or Field.repeat may now follow a variable-size field (#38, @samoht)
• Codec.encode no longer requires an ?env for a codec whose only parameters are decode-side outputs (a field with an Action.assign into a Param.output). Output params are never read when encoding, so demanding an env raised Invalid_argument spuriously, and an output-param sub-codec embedded as a field could not be encoded at all (#95, @samoht)
• Codec.size_of_value now sizes a Field.repeat with a dynamic budget, a Wire.casetype field, and a packed bitfield (wrapped by Wire.bit or an enum / map) correctly. The first two were under-counted (so Codec.encode overran the buffer) and the bitfield over-counted (so encode raised a spurious Invalid_argument) (#72, #78, #79, @samoht)
• Codec.encode now raises Invalid_argument when the writer emits fewer bytes than size_of_value promised, instead of shipping a value with uninitialised trailing bytes (#62, @samoht)
• Codec.encode into a too-small buffer now fails with a precise byte count instead of writing past the end (#61, @samoht)
• Field.optional / Field.optional_or with a dynamic gate now encode from the value (the gate selects the decoded value or the default on decode); optional raises Invalid_argument rather than writing a phantom byte or overrunning the buffer when the gate and value disagree (#58, @samoht)
• Field.optional / Field.optional_or predicates that use bitwise / shift / mod operators are no longer silently treated as always-true, and Field.ref on an optional field now reads the decoded value instead of 0 (#48, @samoht)
• Decoding an all_zeros field that contains a non-zero byte now returns a Constraint_failed error instead of raising (#60, @samoht)
• Wire.to_string on a Wire.nested ~size:n field now zero-pads to n bytes when the inner writes fewer, so it agrees with Wire.of_string (#60, @samoht)
• Wire.to_string on a codec field whose inner ends in all_bytes / rest_bytes / all_zeros no longer appends a 4 KB scratch tail; the size is computed from the value (#54, @samoht)
• Codec.encode / Codec.raw_encode accept ?env:Param.env, like Codec.decode. Encoding a parametric codec with a missing param binding now raises Invalid_argument naming the param instead of writing zero-sized regions (#53, @samoht)
• Fix C stub generation for schema names with two or more leading capitals (e.g. IPv4, EP_Header) (#36, @samoht)

0.9.0

Initial release.