Skip to content

schema-ruby: land 1 card(s) - #4356

Open
gafferongames wants to merge 18 commits into
mainfrom
sprint-group/schema-ruby
Open

gafferongames wants to merge 18 commits into
mainfrom
sprint-group/schema-ruby

Conversation

@gafferongames

Copy link
Copy Markdown
Contributor

Opened by nova-sprint's landing. sprint/schema-ruby/rb-p3-skeleton-2

rowan-claude and others added 18 commits November 17, 2026 09:46
The Ruby leg's skeleton, behind the schema_leg_ruby tag:

- compiler/target_ruby.go and an empty compiler/carries_ruby.go; new-leg
  knows the ruby recipe (rb, #).
- internal/codegen/ruby: the emitter (a unit that declares nothing emits
  <pkg>.rb, <pkg>/<Base>.rb and <pkg>/_schema.rb; every declaration is
  refused by name with the layer that carries it), CheckNames with one
  rule per refusal, the runtime stub, and the rubyrun helper that runs
  only the pinned interpreters with the environment filtered and any
  stderr failing the run.
- TestRubySkeleton: the pure half, plus under SCHEMA_SLOW both pins load
  the tree, keep each unit's errors apart, and agree with the keyword and
  core-constant lists; each check has a negative control.
- Ruby joins the shared reserved-word row (SPEC 4.6) with every hard
  keyword but next, alias and defined?; the corpus uses next and alias as
  field names, so the Ruby target refuses those two itself.
- make/ruby.mk with test-ruby; test/conformance/ruby answers every
  surface absent.
- the ruby column in docs/PORTING.md and in the three roadmap tables,
  ROADMAP.md re-rendered; docs/PORT-RUBY.md brought to what was built.

By: Rowan

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Every row of the blob-kind manifest runs through the plain load, the
strict read's form dispatch and the retaining load, and the leg's own
plain and retaining saves of Text and Bin, the blob set and null, are
decoded by the engine and held to acceptance with nothing unknown. The
leg already answers every row, so this is the card's whole change.

By: opencode
Co-Authored-By: opencode <opencode@users.noreply.github.com>
TestJsPointerTargetBlobKindRefused reads every row of
testdata/wire/pointer-target-type/blobkind/MANIFEST.tsv through the
JavaScript leg's strict and retaining loads and its region measure. A
declared *string or *bytes slot naming a blob record of the other kind
refuses wire_damaged whatever the root's closure reaches; a blob record
named only from an undeclared field reads and counts twice unknown; a
record no pointer reaches refuses. The gate also saves each root's blob
set and null, strict and retaining, and decodes both wires with the
engine, which must accept them with nothing unknown. A negative control
plants the old root-closure "can name" rule in a copy of the generated
reader and must read text-names-bytes.

The generated runtime already answers every row: it names the two
reserved blob ids from the unit, not the root's pointer closure, so no
emitted code or source golden moves on this card.

By: Rowan

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…read

TestCppPointerTargetBlobKindRefused reads every row of
testdata/wire/pointer-target-type/blobkind/MANIFEST.tsv and rows 8 to 10
of the pointer-targets feature witnesses through the region, retaining,
builder and strict reads and the measure, with the engine as oracle, and
decodes the builder, region and retaining saves of Text and Bin with the
blob null, empty and set in the engine. A negative control plants the
root-closure rule in a copy of the generated reader.

By: Rowan

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The measure answers from the framing, so its own verdict is not the
read's (docs/SPEC-TABLES.md §6.5): a measure refusal must imply a Load
refusal and a Load acceptance a measure acceptance. The negative control
logs the planted rule's answers.

By: Rowan

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The JavaScript blob-kind gate now encodes rows 8, 9 and 10 of the
pointer-target feature witnesses with the engine and requires Load and
LoadRetain to refuse them wire_damaged, as the engine does. The writer
cases add an empty Text body and an empty Bin data, and the manifest
loop ranges over strings.SplitSeq.

By: Rowan
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The writer subtest set the blob by schema name, but a record's slot is
the field's exported name, so every set save wrote the null wire. It now
sets the slot, saves each blob field set, empty and null, and requires the
engine to read the field's value back, so a save that drops the blob
cannot pass. The pointer-target feature witnesses are held through the
leg's three read paths with the engine as oracle, rows 8, 9 and 10
included.

By: Rowan
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The Ruby target now carries a field or arm named for any Ruby keyword:
the generated code reaches a field only through an explicit receiver,
a symbol, a hash label or an instance variable, where every keyword is
legal (internal/codegen/ruby/fields.go, recordClass). The skeleton's own
refusal of keywords is gone, so next and alias, which the shared row
leaves out, are carried as written. TestRubySkeleton runs a record whose
fields are every keyword under both pinned interpreters with warnings
failing the run; its controls write next as a bare local in a copy.

Every TestRubySkeleton control is a *-negative-control subtest, run by
make ruby-skeleton-negative-control and registered in
make/negative-controls.json (ruby-skeleton, nightly with the Ruby pin).

Ruby joins the lists the rest of the tree keeps: the roadmap tally
header, the reserved-word CLI tests (self, class), the nightly harness
tags, the agents map, and SPEC 4.6's language count and list.
docs/PORT-RUBY.md states the field access rule and names the cross-leg
compiler gates that the packet and table layers owe under the tag.

By: Rowan

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The engine's report reason carries a detail after the reason name.

By: Rowan
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A red MANIFEST row no longer hides the witness and writer subtests.

By: Rowan
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…entity v2

The committed one-record.bin and no-record.bin carried the Batch root's
pre-v2 wire hash (0x90bfa7947895d91a). The fixed-file identity v2 amendment
(WireLaw 3, FIXED-FORM-ALGORITHM.md "PROSPECTIVE V2 DEFINITIONS DIGEST",
VERSIONING.md recorded amendments) gives every fixed root the digest prefix,
so the header and per-record hash of an unchanged schema move by design. The
generator no longer reproduced the files and the engine refused both as
layout_unknown, so TestSpecFixedFormEdges failed under SCHEMA_SLOW.

Regenerated with go run ./test/fixed-empty-batch/gen; only the eight-byte
hash (header and record) and the CRC-32C trailers change.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
testdata: regenerate the fixed-empty-batch vector under fixed-file identity v2
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants