Repository navigation
Physics Hinge World Selected Reference
이 문서는 코드 전 선택이다. 실제 World 계산·build·types·화면 검증은 실행하지 않았다. Root 준비 계약, API, controls와 함께 사용한다. 기존 완료 보고14개와 부모93개를 다시 실행하지 않는다. 실제 native 채택 기준과 reference authority/expectedRefusal/initial 문법은 선택 문법 부록에 정리했다. protocol controls26개도 선택했다. 실제 채택 provider와 이 선택 문법은 새 작업 공간의 before-code addendum에서 소스에 연결한다. 문서 선택이 실제 source freeze를 대신하지 않는다.
World 요청의 정확한 top-level 키는 {op,config,sourceState,dt,commands,commandsRawHex,outcome,attempt,output,error,rawFrames,negativeZeroPaths}다. op=world-step, outcome=returned 또는 refused. config는 원래 고정4몸체/2관절 rest 입력이다. sourceState는 앞 단계에서 독립 검증한 accepted public state다. dt=1/120, commands는 이 호출의 입력이다. output은 정확한 {state,diagnostics}이며 state는 실제 public return, diagnostics는 accepted diagnostic이다. 거절 때 output=null이다. attempt는 실제 copied last-complete-stage 값이다. error는 원래 알려진 PhysicsError의 safe record 또는 null이다. 예상 결과나 candidate row/basis/optimum seed를 넣지 않는다.
rawFrames는 정확한 six keys {source,angular,motor,free,postContact,projected}다. 도달한 full frame은 고정 body order의416B Float64LE raw hex이며 미도달 frame은 null이다. 몸체 순서는 pairA0,pairB0,pairA1,pairB1, scalar 순서는 p3/v3/q4/omega3이다. Pair projection208B raw는 선택 pairProgress.projectedRawHex에 보존한다. 전체 projected frame은 두 pair가 완료해야 존재한다. commandsRawHex는 원래 angular command 순서의24B씩이다. JSON 값과 raw 대응을 검사하며 zero 부호는 raw가 기준이다.
negativeZeroPaths는 정확한 {sourceState,commands,attempt,output}의 네 dense string arrays다. 각 값은 해당 own-data root에 상대적인 JSON Pointer다. Null root는[]다. 중복·미존재·숫자0이 아닌 경로는 REFERENCE_INPUT이다. Raw scalar와 겹치면 부호가 같아야 한다. Initial 요청은 정확한 {op,config,sourceState,sourceRawHex,negativeZeroPaths}이며 op=initial이다. 동일 four roots를 유지하되 sourceState 외에는[]다. Initial은 graph별 한 번, step 요청과 별도이며 motor/free/projection 호출이 없다.
새 outer reply는 {ok,resolved,executionFailure,referenceStatus,metrics,reference,violations,error}다. Status는 RETURN/PRODUCER_REFUSAL/MISMATCH/UNRESOLVED다. Protocol 오류 리터럴은 REFERENCE_JSON / hinge world reference line must be valid JSON, REFERENCE_INPUT / hinge world reference input is invalid다. 두 경우 outer=false/false/true, UNRESOLVED, metrics/reference=null, violations=[], safe error={code,message}다. 실제 숫자의 gate 실패는 resolved MISMATCH이며 malformed 입력으로 바꾸지 않는다. Nominal strict cap 거절은 정확히 예측해도 실제 World NO-GO다.
Root selects no new reduced parent proof. reference.pairReplies is a fixed two-element array in hinge0/hinge1 order. Every reached pair reply is the complete unchanged actual #1349 independent reply, including every nested coefficient, matrix, inverse, exact Fraction proof, actual rounded certificate, physical observation, scalar disclosure and zero convention returned by that provider. Do not replace any parent field by a hash, compact certificate or reconstructed subset. Parent9-key physical success is {ok,resolved,executionFailure,metrics,violations,reference,certificate,transcendentals,zeroAuthority}. This is NOT the new #1179 outer8-key envelope. Parent safe REFERENCE_INPUT/REFERENCE_SYSTEM reply shapes likewise remain original; do not pad them into a fictitious9-key success or convert them into a World outer error without retaining the original reply.
Exact source pointers, read only: examples/physics-rigid-hinge-velocity-reference/reference.py:602 physical success return; lines644–648 original safe input/system refusals. Its evaluate_physical is called from independently authored POSTCONTACT input with actual pair projection output supplied only for separate comparison. reference.reference/certificate nested schemas and units are inherited unchanged from that exact file and frozen provider graph. The current #1179 outer envelope is defined by examples/physics-rigid-hinge-motor-reference/reference.py:773–787; only that outer8-key pattern is reused by the World wrapper. Exact source identity must be pinned at final source freeze, not inferred from line numbers.
Parent prediction/default actual flags cannot count as actual pair qualification. World metrics adds pairActualCompared:[boolean,boolean], recording whether each complete actual projection result/raw208B was supplied. Null pairReplies[i] means no parent evaluation completed. Full prediction proof without actual projection may be retained only with pairActualCompared=false; preserve any earlier safe projection input/error in attempt. A parent reply with no actual supplied may be retained only as an explicitly unreached/prediction proof; pairActualCompared=false. Successful World RETURN requires both true and both parent resolved/qualified. Never treat a parent prediction as actual success or omit a reached parent's full proof because a later report/cap fails.
Minimal exact outer success/refusal keys remain {ok,resolved,executionFailure,referenceStatus,metrics,reference,violations,error}. Selected World reference top-level keys are the nine {stage,authority,parsedActions,frames,rawFrames,pairs,contactChecks,calls,expectedRefusal} PLUS pairReplies. No compactParentProof field. pairs contains only separately authored World motor/report/work/phase diagnostics, not a duplicate full parent reply. Its selected exact keys are {jointId,motor,jointReport,bodyCorrections,energyByPhase,work,reportCap,phaseMetrics}. motor/report/corrections/cap keys inherit qualified #1179 types; energy adds postContactJ and work adds contactDeltaKJ with POSTCONTACT projection baseline, as original proposal. phaseMetrics has six phase labels and raw/normalized independent metric records; the full POSTCONTACT/projected parent observations stay in pairReplies unchanged even if a World phase metric duplicates an individual scalar. No matrices/Fraction proof repeated in pairs.
World metrics exact top-level keys become {actualCompared,pairActualCompared,checks,frameErrors,pairErrors,contactChecks,reportErrors,workErrors,precision}. Fixed two-pair errors/order and six phase labels remain. checks compares complete actual input/output/attempt/raw semantics; unreached values null, not fabricated true. Parent unsafe/system refusal with preserved safe native error is classified according to original parent error contract, never rewritten as protocol bad input merely because its reply uses a different safe shape. Any unmatched/unresolved parent protocol/resource error produces outer EXECUTION_FAILURE; valid resolved physical/cap disagreement stays World NO-GO.
- SOURCE: restore binary source13fields/body and remembered accepted report. Validate literal rest config separately. Source cursor n/time=n*dt is provider envelope authority; attempted n+1/time=(n+1)*dt remains World transaction metadata.
- ANGULAR: independently execute original angular impulse+torque*dt/inverse-inertia leaf-rounded operations in original global body order, including zero additions. Keep input−0; output−0 follows each actual operation.
- MOTOR: independently execute original per-joint motor on remembered angle, original signed endpoint rule, scalar inertia and clamp. No raw geometry substitution. Separate j/sBefore/sAfter and signed midpoint/spin work.
- FREE: independently original omega/movement bounds, scalar quaternion advance then original linear operations and p advance. Scalar sin/cos/hypot seeds are binary calculations, not Decimal transcendental certificates.
- POSTCONTACT: independently inspect all four cross-group sphere pairs in original pair order at all8 passes. The named successful profile admits no detector hit, including exact touch. This proves expected POSTCONTACT=expectedFREE raw; candidate detector null alone is not proof. Connected pairs remain excluded. If genuine contact/touch occurs, retain original native contact work and classify named physical NO-GO; do not seed expected POSTCONTACT from actual or expand into an independent contact solver campaign.
- PROJECTED: independently compute each POSTCONTACT pose, centered momentum targets/rows/H/x0, Fraction branch optimum and Decimal80/120 raw+normalized measurements using unchanged parent functions. Targets/report baseline are POSTCONTACT, never FREE or prior zero momentum. Independently derive J/K, target report memory, cap and signed work. Actual rounded output receives a separate certificate and geometry/budget comparison. No candidate rows/axis/basis/u/parent-output seed determines expected optimum.
Previous accepted candidate state is admissible per-step ORIGINAL input only after the prior independent accepted-step gates pass. It is not an independently expected global trajectory cache. Each ask recomputes expected work cold, no expected answer cache, no parent14/93 replay. Preserve independent per-step and long-run COM/P/L/work gates already proposed; raw/norm1e-12, algebra1e-25, scalar1e-12, parent original SI gates and strict .01 cap remain unchanged.
Original order: disposed→dt→MAX_STEPS→command parse→EACH joint axis admission→angular/motor/free bounds→contact passes→pair0 projection/report→pair1 projection/report→finite diagnostics/original result validation→copied observer→strict hinge0/hinge1 caps→single commit. Original result validation is within the selected calculation catch, before observer/caps/commit. Registered observer at most1 and outside calculation catch. Observer throw wins over later cap and propagates exact received identity. Admission callback0. Known earlier errors preserve exact class/code/message and completed stage; no conversion to CAP.
Unknown received values propagate unchanged without new reflection/serialization; the inherited typed-parent proxy boundary remains a limit on this guarantee. Harness records only safe phase, known-vs-unknown flag, counts and previously copied safe data; identity is asserted in memory by a separately counted control. No fabricated known error/reference request. If safe producer outcome cannot be obtained, preserve EXECUTION_FAILURE and stop. Reference child exception/no reply/protocol/unresolved/storage/IPC failures are EXECUTION_FAILURE. Resolved actual geometric/work/momentum/report disagreement, nominal original numeric refusal and any cross-group contact are physical World NO-GO. Stop source at first failure, then later graphs execute the same frozen prefix once; no tuning.
Persist request exact JSONL bytes synchronously BEFORE sending. Capture stdout/stderr raw chunks before parsing; drain pipes to child close before final receipt. Keep first rejection, add secondary close/storage failures separately. Exit receipt has code AND signal, exact stdin/stdout/stderr byte counts/hashes, complete-line counts, trailing fragment bytes and paths. Stderr counts use raw bytes, not reconstructed UTF8 length. A nonzero exit/missing reply is not an expected numeric refusal. Preserve malformed stdout, partial LF/JSON and first failed request; never synthesize its reply. Accepted diagnostics/cache/cursor remain prior values on all refusals; attempt and accepted traces remain distinct.
Selected receipt top keys {pid,exitCode,signal,stdin,stdout,stderr,completedReplies,pendingRequestIndex,firstFailure,secondaryFailures}. pid is positiveinteger or null when spawn never completed; exitCode integer or null; signal string or null, both copied from actual close arguments. Each stream exact {path,bytes,sha256,completeLines,trailingBytes,completeEOF}; count bytes on raw chunks, hash exact stored bytes, never re-encode decoded stderr. stdin.completeEOF means the writer ended/closed its pipe, NOT proof the child consumed every byte. stdout/stderr.completeEOF means their actual pipe close was reached after drain. completeLines counts LF-delimited physical lines; trailingBytes is the exact non-LF suffix length. completedReplies counts actually parsed valid replies, not stdout lines; pendingRequestIndex is null or original request ordinal awaiting a reply. Protocol malformedJSON controls remain actual requests/replies and count separately.
firstFailure is null or safe wrapper-generated {phase,code,message} with fixed own-data literals; secondaryFailures is an array of the same safe wrapper-generated form. Do not inspect arbitrary observer/model thrown value to produce these fields; native unknown identity proof remains separate safe boolean/phase control evidence. First rejection stays primary; later close/logging/storage failures cannot replace it. Hash/path receipts are written only after pipes drain or explicitly marked incomplete. Exact request bytes are durable before sending, output raw logging precedes line parsing. Failure literals below and Root bounded shutdown govern this receipt. Actual implementation/provenance are frozen before execution.
Fixed phase order is source,angular,motor,free,postContact,projected. Fixed pair array length2 is hinge0,hinge1. JSON objects have only listed own keys; arrays are dense. E means finite NONNEGATIVE maximum absolute error, or null iff its corresponding actual comparison was not reached. B means boolean or null iff unreached. Never use Number(null), bool-as-number, missing/sparse fields or finite coercion. A reached exact zero comparison is numeric0; unreached is null. Complete RETURN requires all actual-frame/error/report/work fields reached and both parentActualQualified=true. Producer refusal permits only its honest reachability null mask. Outer unresolved is EXECUTION_FAILURE before any physical numeric gate.
metrics exact keys remain {actualCompared,pairActualCompared,checks,frameErrors,pairErrors,contactChecks,reportErrors,workErrors,precision}. actualCompared is boolean; pairActualCompared is two booleans, false for no complete actual projection. No parent prediction defaults qualify actual comparison. Unknown producer cannot supply a fabricated metrics record.
checks exact keys:
{sourceRawMatchesInput,commandsRawMatchesInput,negativeZeroPathsMatch,sourceCursorRetained,attemptCursorCorrect,parsedActionsMatch,heldFrameBitsMatch,namedZeroSignsMatch,outputAttemptCorrespondence,frameReachabilityMatch,pairProgressMatch,contactSequenceMatch,callsMatch,finiteSignedWork,nonnegativeEnergiesAndObjective,actualErrorMatch,actualOutcomeMatch,pairs}.
The first17 values are B. pairs is fixed2 records, each exact {acceptedReportAuthorityMatch,motorMetadataMatch,motorBranchMatch,jointReportIdentityMatch,targetAngleAuthorityMatch,reportBaselineIsPostContact,capDecisionMatch,parentActualQualified} with B values. Complete CAP has outputAttemptCorrespondence=null because output is absent, not fabricatedtrue; all actually reached pair/report comparisons remain required. Full accepted/hidden-store rollback is verified by native snapshot/control evidence, NOT by a reference check that only sees the same sourceState twice. Do not add a tautological acceptedRetentionMatch flag to this wire.
frameErrors is exact six-phase object, each value null or {positionM,quaternion,velocityMps,angularVelocityRadps,metadataMatch}. Four numeric E values have m/dimensionless/ms^-1/rads^-1 units. metadataMatch is B for a reached frame. Maximum is over all four raw-restored bodies/components. Metadata comparison includes exact immutable IDs/mass/inertia/material/kind/shape, not a re-parsed moving constructor.
pairErrors is fixed2 six-phase objects; each phase null or {positionM,angleRad,quaternion,velocityMps,angularVelocityRadps,linearMomentumKgMps,angularMomentumKgM2ps,energyJ} with eight E values. These are grouped errors between actual published phase diagnostics and independent measurements, using the qualified #1179 grouping at reference.py615–630 (position anchors/gap/COM, angle raw/axis/limit, unitq, anchor velocities, relativeomega, P, orbital+spin+totalL, KE). Parent raw/normalized POSTCONTACT/projected proof remains independently gated in full pairReplies; it is not replaced by these scalar maxima. Unreached published PairDiagnostic can leave pairErrors null even when pairProgress contains a returned projection; the parent reply still qualifies that actual projection.
reportErrors is fixed2, each null or exact {angleRad,linearImpulseNs,angularImpulseNms,motorImpulseNms,bodyLinearImpulseNs,bodyAngularImpulseNms,angularReactionNormNms,motorSpeedBeforeRadps,motorSpeedAfterRadps}. All E, units rad/Ns/Nms/Nms/Ns/Nms/Nms/rads^-1/rads^-1. Null pair record is allowed only before both motor and report comparison are reached. After actual motor returns, motorImpulseNms/motorSpeedBeforeRadps/motorSpeedAfterRadps are finite errors and the six unreturned report fields remain null. When actual report returns, all nine errors are finite; motorImpulseNms also checks the actual report's copied j against the same original motor j. Reuse qualified report comparisons at #1179 reference.py735–754, rebasing ONLY correction baseline to independently derived POSTCONTACT. Speed errors compare actual motor diagnostic to own original-motor derived before/after speeds; branch/metadata authority is Boolean in checks.pairs. A returned motor diagnostic never evades speed validation when a later projection/report fails.
workErrors is fixed2, each null or exact {externalAngularByBodyJ,externalAngularTotalJ,motorMidpointJ,motorSpinMidpointJ,motorResidualJ,freeDeltaKJ,contactDeltaKJ,projectionDeltaKJ,projectionObjectiveJ,energyByPhaseJ,motorBalanceResidualJ}. All E in J; first is max over original2-element per-body work. energyByPhaseJ is max over all six phase energy values. motorBalanceResidualJ is abs(actualSpinMidpoint-actualMotorMidpoint), not abs(signedWork) and not total conservation. The original signed work remains untouched in actual full payload and reference.pairs. Objective gate1e-10J; other work error gates1e-12J. Null until actual complete work record exists; no forged partial zero record.
precision exact {algebraMaxDifferenceByUnit,scalarRepeatMaxDifferenceRad,algebraResolved,scalarResolved}. algebraMaxDifferenceByUnit exact {positionM,quaternion,velocityMps,angularVelocityRadps,linearMomentumKgMps,angularMomentumKgM2ps,energyJ,impulseNs,angularImpulseNms}; each E over actually independently evaluated quantities at80/120, null only if none reached in that unit. scalarRepeatMaxDifferenceRad is E, with independently repeated scalar angle seeds; no Decimal transcendental claim. algebraResolved/scalarResolved are booleans. A null precision unit is not numeric0 and cannot qualify a reached quantity; both complete flags true required before physical gates. Parent precision values/certificates remain full original pairReplies as well.
contactChecks metrics exact {passes,final}. passes is dense length32, final dense length4; each slot null until its actual check was reached. Each reached pass slot exact {pass,pairIndex,aId,bId,positionsMatch,hitMatch,noHit,separationRawM,separationNormalizedM,precisionMaxDifferenceM}. pass/pairIndex are original integer indices; three flags B; separationRawM/separationNormalizedM are FINITE SIGNED distances-minus-radii in m, not E; precisionMaxDifferenceM is E. IDs exact. Final slots have the same keys except pass is null (fresh final check, not ninth original pass). NoHit must be true in successful named profile including touch refusal; signed separation must support it independently in both modes. Geometry branch/touch differences are recorded and are physical disagreement; never assume candidatehit=null proves separation. No normal/point errors are needed for successful no-hit checks; actual hit/point/normal remains complete in request. A hit terminates this no-contact profile, not an expansion into a new contact law.
Error text in outer violations retains fixed field/phase names plus known pair IDs. Known exact producer code/message remains preserved separately in request/error/expectedRefusal. Booleanfalse or E above its frozen SI gate yields resolved numerical MISMATCH; schema/nonfinite/missing/unreached inconsistency yields REFERENCE_INPUT or unresolved EXECUTION_FAILURE as selected by actual boundary. Do not turn genuine numerical values into malformed-request errors merely because they fail a gate.
The following tuples are WRAPPER-generated fixed data, not reflected messages from unknown native/observer payloads. They appear only in childreceipt firstFailure/secondaryFailures or corresponding safe consumer resource/IPC evidence. They do not become producer PhysicsError, parent refusal, expected CAP or numerical GO.
| phase | code | exact message |
|---|---|---|
| spawn | IPC_SPAWN_FAILED | hinge World reference process could not start |
| request | IPC_REQUEST_WRITE_FAILED | hinge World reference request could not be written |
| request | REQUEST_BYTE_LIMIT | hinge World reference request byte limit exceeded |
| stdout | IPC_STDOUT_READ_FAILED | hinge World reference stdout could not be read |
| stdout | REPLY_BYTE_LIMIT | hinge World reference reply byte limit exceeded |
| stdout | IPC_STDOUT_PERSIST_FAILED | hinge World reference stdout could not be stored |
| stderr | IPC_STDERR_READ_FAILED | hinge World reference stderr could not be read |
| stderr | IPC_STDERR_PERSIST_FAILED | hinge World reference stderr could not be stored |
| request | IPC_STDIN_PERSIST_FAILED | hinge World reference request could not be stored |
| reply | IPC_REPLY_JSON_INVALID | hinge World reference reply is not valid JSON |
| reply | IPC_REPLY_SCHEMA_INVALID | hinge World reference reply schema is invalid |
| reply | IPC_REPLY_UNSOLICITED | hinge World reference returned an unsolicited reply |
| reply | IPC_REPLY_MISSING | hinge World reference exited before its reply |
| close | IPC_EXIT_NONZERO | hinge World reference process exited unsuccessfully |
| close | IPC_EXIT_SIGNAL | hinge World reference process terminated by signal |
| close | IPC_CLOSE_RECEIPT_FAILED | hinge World reference close receipt could not be stored |
| proof | PROOF_BYTE_LIMIT | hinge World proof byte limit exceeded |
| proof | PROOF_PERSIST_FAILED | hinge World proof could not be stored |
| parity | PROOF_BYTES_DIFFER | hinge World proof bytes differ between graphs |
| producer | NATIVE_UNKNOWN_THROWN | hinge World producer threw an unknown value |
| observer | OBSERVER_UNKNOWN_THROWN | hinge World observer threw an unknown value |
For child close with pending request, IPC_REPLY_MISSING is primary unless an earlier captured failure already exists; nonzero/signal is retained as secondary and raw exitCode/signal always remain actual. A signal with no pending request selects IPC_EXIT_SIGNAL before IPC_EXIT_NONZERO. Parsing a malformed reply records its raw bytes first; first failure is never replaced by laterclose/storage failure. Source reference protocol JSON/INPUT errors remain exact outer {code,message} literals in A2 and are not confused with child malformed-output errors. Known numeric World refusal is not an IPC failure tuple.
Overflow policy: no new request, World call, helper, comparison campaign or graph continuation after overflow. Close stdin, terminate the current child under Root-selected bounded shutdown policy, drain/preserve actually captured stdout/stderr/raw/request prefix and pipe EOF/fragment facts, mark incomplete proof truthfully. Return GRAPH_NO_COMPLETE_PROOF disposition / executionFailure=true, not physical World NO-GO and not complete full trace GO. Do not fabricate a response or alter a producer/model value to fit. Any emitted overflow error envelope has outerUNRESOLVED/error=null with literal violation hinge World reference byte limit exceeded; raw IPC remains the actual received prefix. Emergency8MiB retains actual divergence/failure prefix within its own fixed ceiling. If termination prevents full drain, completeEOF=false and trailing captured bytes/hash are explicit. Parent system/rank/precision/resource algorithm remains unchanged.
Main1,024MiB + fixed64MiB + emergency8MiB =1,096MiB다. 요청256KiB, 전체 reply512KiB, compact receipt4KiB다. SOURCE1,200단계의 IPC 계획 상한900MiB, receipt4.6875MiB, global raw2,995,200B와 initial416B 및 실패 여유를 main에 포함한다. Controls<=32MiB는 fixed에 포함한다. 이 값은 실제 World 크기의 측정이나 모든 입력에 대한 보장이 아니다. 중간292/584MiB 계획은 선택 전에 대체했다. 물리 입력·cap·수치 gate는 바꾸지 않았다.
전체 실제 stdin/stdout/raw를 저장한다. Canonical receipt는 byte ranges/hash/호출/판정만 연결한다. Summary는 counts/extrema/first-failure locator만 유지한다. 전체 observations 배열을 중복하지 않는다. Repeat/compiled는 실제 호출과 전체 byte 비교를 한 뒤에만 SOURCE bytes의 검증 alias를 저장한다. 차이가 있으면 actual divergence prefix를 emergency에 보존하고 중단한다. 세 graph 전체 IPC를 중복 저장하는2,700MiB 계획은 선택한 한도 밖이다. 압축률이나 alias를 SOURCE budget에 가정하지 않는다. Runtime 저장 한도는 Git archive 추가 복사나 부모 Fraction 내부 메모리 상한이 아니다. Archive bytes는 별도 기록한다.
Overflow 뒤 새 요청·World 호출은0회다. Stdin 종료와 SIGTERM 후 실제 close를 최대5,000ms 수집한다. 종료되지 않으면 SIGKILL 한 번과 추가5,000ms drain을 적용한다. 종료되지 않은 pipe의 completeEOF=false, 실제 captured prefix/hash/trailing fragment/signal을 보존한다. Complete drain이 없으면 GRAPH_NO_COMPLETE_PROOF다. 저장/transport 오류를 수치 World NO-GO로 표시하지 않는다. 각 transport buffer8MiB와 emergency disk8MiB는 별도다. 전체 payload 예정 hash와 실제 저장 prefix hash를 혼동하지 않는다. 한도를 올려 성공 prefix를 다시 실행하지 않는다.
Full main GO의 graph별 계획 수: actual World step1,200, initial+accepted state1,201, motor/projection/report 각각2,400, 몸체 angular/free 각각4,800, contact passes9,600, producer detector38,400, 독립 final checks4,800, global full raw frame7,201, pair phase blocks14,400이다. Graph3개면 main3,600회·pair projection7,200회·global raw21,603개다. Controls·실패·protocol 요청·부모 precision/expected/actual solve는 별도 actual 수를 기록한다. 이 수는 완료된 실행 결과가 아니다. 실패하면 attempted n+1와 accepted n을 구분하고 도달한 실제 prefix만 기록한다.
각 stream에 bytes/SHA/LF/ordinal/offset/trailing/completeEOF를 기록한다. Source/dist closure·bundle·emitted·build/types·owner 출처는 canonical parity 밖의 고정 provenance다. 실패한 요청이나 원래 stderr를 재구성하지 않는다. Parent schema/solver/허용값을 바꾸지 않는다.