Files
prometeu-studio/tools/pbx-conformance/README.md
T
bquarkz c410e35001 feat(pbs): add mutable lists and stable snapshot cursors
Implement SDK list<T> and Cursor<T> with trusted compiler bindings and generic storage publication across IR and PBX. Pin the runtime provider and include specs, conformance coverage, installed examples, and DSC-0090 lessons.
2026-10-23 00:46:39 +01:00

267 lines
16 KiB
Markdown

# Studio PBX producer conformance
Owner: compiler/general. This harness uses the production runtime loader and VM;
it does not implement a Java approximation of runtime admission.
From the Studio root:
```sh
./gradlew :prometeu-compiler:prometeu-build-pipeline:test --tests '*LinkedRegistryTest' --tests '*PbxRegistryConformanceTest' --tests '*TypedCommonPipelineTest' --tests '*RuntimeCatalogPipelineTest'
./gradlew :prometeu-architecture-tests:test --tests '*PbsRuntimeAlignmentTest' --tests '*PbsStructuredValuesTest' --tests '*PbsClosuresTest' --tests '*PbsGenericContractsTest'
python3 tools/pbx-conformance/run.py --runtime ../runtime
```
The Java tests generate PBX in the build-pipeline module's
`build/registry-conformance`. `run.py` creates a temporary Cargo project, copies
the runtime lockfile and references its bytecode/hal/vm/system crates by path.
It never modifies the runtime checkout. Add `--online` for an uncached machine;
by default Cargo uses only cached dependencies. Java 25, Rust and the runtime
checkout's dependencies are required.
The output identifies runtime revision, OS/architecture and load configuration.
Positive fixtures distinguish loading from execution. The vec2 signature fixture
is deliberately load-only because its entry requires arguments. Native binding
uses SYSTEM for the positive case and NONE for missing-capability rejection;
all other cases use NONE. Resource tests explicitly lower default limits.
Negative fixtures mutate Java-produced bytes independently of the writer to
exercise malformed input that the writer correctly refuses to construct.
Initialization JSON comes from `prometeu_system::CrashReport`, and is checked
against the runtime's published headless schema. Saved JSON files remain under
the generated fixture directory. Studio process tests replay selected observed
payloads from `prometeu-app/src/test/resources/runtime` and assert unchanged MCP
projection, including null executed PC and nested verifier context.
## Known upstream schema defect
At runtime `f7d7c167531bb662a1af9d73fc326bb476bf8aa7`, the LoadReason Rust enum
includes LifecycleRole, LifecycleSignature, DuplicateLifecycleRole and
NonFiniteValue, absent from the published JSON schema enum. This corpus
observes LifecycleRole and NonFiniteValue. The harness prints an explicit
`UPSTREAM SCHEMA GAP` for those two reasons and verifies that adding exactly
the missing enum value makes the original payload valid. It never alters the
payload or reports those cases as conforming to the unmodified schema. Any
other schema failure remains fatal. Track the runtime correction in AGD75;
remove the exception once the schema is corrected.
The original registry/catalog fixtures qualify common backend production. The
PBS fixtures below additionally qualify source lowering; this harness does not
claim installed-runtime deployment or memcard persistence.
## Runtime catalog consumer fixtures
The catalog corpus adds common-IR Random calls with an aliased handle, a discarded
first result, mixed next/nextf/range calls and a subsequent next. The VM checks the
independent seed-42 handoff vector and identical charged work before/after
optimization. Constants are compared slot-by-slot against runtime materialization,
including exact Float bits and both slots of `vec2.zero`. Pool deduplication may
share a value but cannot omit a group component.
A known zero-argument touch query loads and traps `TRAP_UNAVAILABLE` without a
hardware context. A removed endpoint ID is rejected even after an unreachable
HALT. `color.html_rgba` executes with a String value and rejects a proven null
at load time, matching the runtime verifier's stricter rule despite the
dispatcher's nullable String carrier. These are distinct outcomes. Run the separate provider-export check with
`python3 tools/runtime-catalog/verify.py --runtime ../runtime` to compare the
selected source revision, imported CSV bytes and exporter output as well.
## PBS source and input qualification
`PbsRuntimeAlignmentTest` compiles actual SDK declarations and source programs
through the PBS frontend, production common IR, typed Registry and PBX writer.
Its 114 fixtures execute on the selected runtime: all 48 Pad combinations with
both literal Button arguments and enum locals passed through a typed function,
six Touch queries, numeric boundaries/conversions/faults, builtin groups,
String, fixed multiple returns and named application with observable Y,X effects.
The controlled input platform asserts the exact button/coordinate reader and
one read per public query. Each query is also executed without a platform to
prove TRAP_UNAVAILABLE. The loaded PBS registries are checked for the defaults
of all eight existing value types, including null String, positive Float zero
and complete vec2/pixel groups. No source constructors or new collection API
are introduced to test storage defaults.
The harness requires the complete fixture count. It must not pass by skipping
missing PBS fixtures. Integer arithmetic wrapping is distinct from checked
Int64-to-Int32 narrowing and Float numeric-range faults. The full installed
knowledge gate remains separate and mandatory for SDK resource/producer
integration.
## PBS typed arrays, optional and Struct/contract execution
`PbsStructuredValuesTest` generates 27 mandatory `pbs-structured-*` fixtures.
They cover contextual arrays, nested levels, vacancy/zero, clear/length, aliasing,
int/long bounds, single evaluation, lazy optional fallback, inferred/explicit
optional tuples, native vec2 groups, fields and constructors, overloads and Self,
contract elements/dispatch, imported services/structs/contracts, global file-init,
and the installed example. The runtime harness names every required fixture;
a missing file is an error.
Eighteen additional runs lower GuestBytes or HandleSlots to 0, 1 or 2. They assert
resource faults, bounded accounting, no object publication at zero quota, and
release of session ownership. The integer-storage fixture succeeds with one
handle total, proving that its three integers do not allocate individual boxes.
String session accounting is checked after each structured execution.
The GC fixture compiles all allocations, roots and post-collection reads from PBS.
A test-only driver hook replaces one isolated marker assignment with FRAME_SYNC
and NOPs without changing code offsets; it does not add PBS sync syntax. At the
real runtime safepoint, at least 1,100 discarded objects are reclaimed while
nested references, tuples and strings remain readable. `resource-diagnostics`
is enabled solely for observation. Default production ceilings and GC threshold
are unchanged outside the explicitly lowered quota runs. This is semantic and
resource conformance, not a performance benchmark or installed game qualification.
## PBS first-class functions and closure lifetime
`PbsClosuresTest` adds nine mandatory source/load/execute fixtures: named and
anonymous functions, snapshots, shared objects/arrays, nested function captures,
complete tuple groups and contract receivers, independent environments, Fn fields
and overload selection, imported functions, and the installed example. Each
returns 42. Six additional GuestBytes/HandleSlots runs at limits 0, 1 and 2 check
closure allocation faults and session ownership, bringing structured quota runs
to 24.
The tenth fixture, `closure-frame-lifecycle.pbx`, uses the existing production
lifecycle assembler without bytecode patching. Its first frame stores a closure
capturing an object; the next frame invokes it twice (41, then 42) and drops the
reference. Each of three frames also discards 1,200 closures. The harness asserts
threshold collections, retained closure/object reachability after frame one,
and reclamation after release. The remaining root is the global array.
PBS publishes FRAME_ROOT and the existing guarded wrapper returns normally.
No sync syntax, new frame integration, suspension or coroutine is introduced.
## PBS service constants and runtime math
`PbsRuntimeMathTest` produces 261 independent numeric cases in both ordinary and
optimized bytecode (522 mandatory executions per runtime profile), covering all
26 math/mathf signatures. The checked-in corpus preserves 240 provider scalar
fixture rows and exact contract examples (duplicates removed); provenance is in
`prometeu-architecture-tests/src/test/resources/math/provenance.txt`.
The harness verifies exact bits/faults, selected intrinsic identity and one
retained call per case. Single-step charging is compared with the provider's
opcode costs and independent fixture call cost; full-budget charging must equal
their sum, and optimization must preserve total cycles. Additional cases exercise
pi/tau, literal sin, ordinary int/long service constants, constant dependencies,
and real module imports with service aliases. The installed example also imports
both SDK services, so installedKnowledgeTest covers their packaged resources.
```sh
./gradlew :prometeu-architecture-tests:test --tests '*PbsRuntimeMathTest'
python3 tools/pbx-conformance/run.py --runtime ../runtime --math-only
python3 tools/pbx-conformance/run.py --runtime ../runtime --math-only --release
```
Omit `--math-only` for the full existing corpus. `--release` selects the optimized
Rust runtime harness; it is independent of the PBS optimizer variants included in
every run. Optional `CARGO_TARGET_DIR` can retain build artifacts outside the
runtime checkout. Each invocation prints runtime revision, platform, profile,
features, lockfile and provider catalog/oracle hashes. Local evidence qualifies
macOS/arm64 only; macOS/x86_64 and Linux/x86_64 remain pending.
## PBS explicit Random integration
`PbsRuntimeRandomTest` adds 226 source fixtures plus one real SDK import/alias
fixture. The 48 unchanged provider oracle rows produce 192 exact integer/Float
cases (six seeds, eight draws, two output kinds and two optimizer variants).
The remaining paired fixtures cover mixed/discarded draws, independent objects,
locals/globals/fields/arrays/parameters/results/captures, named/positional effects,
unit/extreme ranges, invalid limits, null and typed-null Dynamic storage, real
rejection followed by acceptance, and collection roots.
Each paired fixture runs with single-instruction budgets and twice with full
budgets, comparing exact result bits or faults, intrinsic charges and total
cycles. Explicit source `as long` conversions are charged separately. The GC
fixture reuses the existing test-only FRAME_SYNC marker hook; all allocations,
roots and post-collection draws are compiled PBS. At least 1,100 discarded Random
objects are reclaimed while array, field and closure references preserve the
same generator. The import fixture uses the production workspace builder and
lifecycle assembler. The focused driver calls file initializers before Probe.main;
it does not claim a new lifecycle API.
```sh
./gradlew :prometeu-architecture-tests:test --tests '*PbsRuntimeRandomTest'
python3 tools/pbx-conformance/run.py --runtime ../runtime --random-only
python3 tools/pbx-conformance/run.py --runtime ../runtime --random-only --release
```
The driver rejects an oracle copy that differs from the selected provider and
prints source/oracle/catalog/lockfile fingerprints and toolchain/profile details.
A complete run includes the Random corpus and keeps it separate from the older
114 input/type fixtures. Rare synthetic 64-attempt acceptance/cap/rollback,
malformed carriers and native provenance are provider tests, not PBS-produced
claims. Qualify those separately with the provider's `random_contract` suite and
`cargo test -p prometeu-vm --lib random` in debug/release. No alternate RNG or
public injection hook is used. Installed knowledge validation remains a separate
gate for packaged SDK/spec/example resources.
## PBS generic contracts
`PbsGenericContractsTest` generates fourteen mandatory source fixtures:
`identity`, `manual`, `multiple`, `tuple`, `optional`, `function`, `contract`,
`native`, `recursive`, `phantom`, `receiver`, `gc`, `installed` and `imports`.
The corpus uses ordinary concrete Trait descriptors on the existing runtime.
It checks several applications on one owner, ordered substitution, composed
multi-slot values, finite recursion, canonical aliases/reexports and distinct
same-named argument declarations. The imported program uses the production
workspace builder and lifecycle assembler.
The generic and handwritten concrete identity fixtures both execute with one
heap handle, establishing that contract adaptation needs no generic wrapper.
Typed absence and implementation index zero are valid. A mismatched Trait call
must be rejected by the verifier or trap with the runtime type check. The GC
fixture retains a contract receiver through a closure while at least 1,100
discarded objects are reclaimed. Reduced GuestBytes, HandleSlots and loader
implementation limits verify refusal and ownership release. Descriptor counts
and PBX byte sizes are printed as evidence, without a performance threshold.
```sh
./gradlew :prometeu-architecture-tests:test --tests '*PbsGenericContractsTest'
python3 tools/pbx-conformance/run.py --runtime ../runtime --generic-contracts-only
```
The full harness also requires these fixtures. The driver uses an external
runtime checkout without editing its sources or changing PBX, ISA or quotas.
## SDK List, shared Cursor and generic storage
Generate the source corpus and the language-neutral storage fixtures, then run:
```sh
./gradlew :prometeu-compiler:prometeu-build-pipeline:test :prometeu-architecture-tests:test :prometeu-app:installedKnowledgeTest
python3 tools/pbx-conformance/run.py --runtime ../runtime --list-cursor-only
```
The full harness also requires this corpus. Missing ordinary, optimized, quota,
GC, import/reexport or installed-JAR fixtures fail explicitly. Two synthetic
common-IR programs execute INDEX_GUARD, ARRAY_CHECK_LENGTH, ARRAY_COPY and
STORAGE_PUBLISH through ordinary and optimized producer paths without PBS types.
List source cases cover every mutator, exact growth, logical versus physical
bounds, int/long arguments (including named application), holes, aliases,
vec2/pixel/tuple/Fn/contract/reference groups and every state of a Cursor that remains on its last node.
An independent producer distinguishes some(none) from an empty Cursor.
Four source-generated mutation fixtures expose explicit before/completed markers.
The harness lowers GuestBytes, HandleSlots, AuxiliaryBytes and TemporaryBytes from
measured setup/success peaks. After each refused attempt it independently observes
all List fields and complete backing groups, including reference generations and
presence tags. Coverage gates require refusals in no-growth publication scratch,
removeAt optional return allocation, candidate allocation, node construction and
operand/frame preparation. Read-only resource diagnostics use non-owning storage
observations outside guest execution; no observation introduces a GC root or write.
The controlled safepoint fixture expects 1,200 nodes/objects initially, one after
traversal, one after repeated next on the final node, and zero after releasing the
Cursor. It also requires no live source/snapshot array. Shared payload fields stay
observable; prefix roots and the final payload obey normal reachability and GC.
The installed fixture is compiled by the isolated installDist JVM using the actual
packaged spec/example/SDK, then loaded and executed by this production VM harness.
Evidence prints compiler/provider source hashes, exact provider revision and dirty
state, toolchain, platform/profile, capabilities, every default VM ceiling and each
explicitly lowered refusal limit. Development runs against a dirty provider are
identified as such; final catalog verification requires its clean selected commit.