gnomonlearn
Docs/RFC-0041 graph install engine conformance

RFC-0041 graph install engine conformance

Versioned draft snapshot

RFCs, reference catalogs, and evidence reflect Spec 3b2e3e7a114fb826b1445d39a48756b8fff629cd. Counts and pending/status notes belong to that snapshot. Read the dated project update for later progress.

Source repositories are private during staging. Links marked “repository access required” need access to the linked repository.

This is a trace and gap record from the pinned draft specification. It does not establish proof or implementation conformance.

Current status — October 18, 2026

The §8 object-failure diagnostic subcheck is closed at tested Gnomon commit e5c2abea9b4b8d12fa25b717f48b17d4a617fd16. A failed object now emits M41020 with manifest namespace/version, object name and kind, while retaining the original error as cause and preserving its cause-specific code. Both continueOnError settings reject before later objects and materialization. The ordinary Roma //graph/tests/graph-module-engine:test_node invocation passed 28 Vitest files and 217 tests on Node v20.19.0 / Darwin arm64, including all seven transaction-test-file cases. The archived actual evidence (Gnomon source: gnomon/tools/roma/evidence/rfc0041-m41020-20261017/README.md; public archive link pending) records the tested revision, real result and independent evidence review.

The approved module rollback-boundary source is adopted at Gnomon fed0662ece0d3ecd6774469085cef261d303ad9e. It removes the implicit batch transaction, adds the explicit isolated client callback, and propagates fatal boundary errors through both entry points. Its new transaction-state-model regressions passed at tested Gnomon commit 4c58e2e7e65666c94bdd1cdef680388011cc4b1c: the ordinary Roma //graph/tests/graph-module-engine:test_node invocation passed all 28 Vitest files and 229 tests. The before/after comparison preserved all 6,367 bound product files and the tool guards. The archived actual evidence (Gnomon source: gnomon/tools/roma/evidence/rfc0041-isolated-transactions-20261017/README.md; public archive link pending) retains the first failed attempt and successful retry. These tests exercise the real client and module transaction service against an in-memory transaction state model, with the installer body stubbed. They do not establish persisted database rollback, caller rollback on real backends, module database commit timing, or per-agent activation integration; custom facade compilation and full framework conformance also remain open.

Separately, the focused ordinary Roma target //runtime/tests/completeness:test_pg_transaction_boundary at Gnomon 091f1f0060fff18cf7168e23a77a88a0f5682beb passed one Vitest file and all five tests against fresh local PostgreSQL 16.15. The five cases took 10.57 seconds (full Vitest duration 11.63 seconds). The five persisted API/backend SQL cases verify owned commit visibility, callback rollback, caller savepoint isolation, caller rollback of successful nested writes, and deferred-foreign-key COMMIT failure followed by reuse of the same writer/backend. An independently pinned observer checks visible rows. The actual scoped evidence (Gnomon source: gnomon/tools/roma/evidence/rfc0041-real-pg-20261017/README.md; public archive link pending) records Root 4328dc/session10051 → c9f827/0, Roma child 55899, and independent acceptance c51d53/0 plus dbe597/0; no skipped or analysis-only result supplies this acceptance.

A later real-PostgreSQL sentinel run at Gnomon 3d83cc412d4a68e46db311ba21de300a54c584bc confirms an outcome-reporting counterexample. The existing Roma //runtime/tests/completeness:test_pg_transaction_boundary target finishes with the original five cases passing and the new sixth case failing (e34a71/session41329 → cf0be4/1, Roma child 22932). The callback catches real SQLSTATE 22012 from division by zero and fulfills. Before line 235 fails on expected rejected, received fulfilled, the test confirms that an independent observer and the same writer see the earlier SQL write rolled back, then successfully commits a fresh transaction through the same backend/writer. The later assertion for an M38061 error is not reached; the raw pg COMMIT command tag is not exposed or directly observed by this test. Independent e671ab/0 joins all 265 artifact bindings and 6,373 unchanged product records; 52c372/0 confirms recorded process/group and endpoint closure. The counterexample archive (Gnomon source: gnomon/tools/roma/evidence/pg-commit-outcome-first-failure-20261017/README.md; public archive link pending) is committed at c103e3f0674350c96285e13083479ad5e785e303. The earlier five-case acceptance retains its tested revision and sentinel-SQL scope; it does not cover this caught-error/fulfilled-callback outcome.

The private five-file repair (local review proposal: .work/pg-commit-outcome-repair-proposal-v1-20261017/README.md; not published; availability pending) proposes forwarding optional actual-driver QueryResult.command metadata and detecting only an actual COMMIT-to-ROLLBACK result, with truthful transaction depth and cleanup tied to that confirmed terminal outcome. The additive public metadata field remains unselected. A separate official adapter-local M38061 repair is adopted and runtime-verified on October 18, 2026 at Gnomon 897cb3902f385fd1d0c4903cffaae6046b810f85. It inspects the actual native COMMIT-to-ROLLBACK result before projection, without adding public query-result metadata. The existing //db/tests/db-postgres:test_node target passes all 17 cases (5f68b6/0, independent cf0ab6/0); its database/config dependencies are mocked. The unchanged //runtime/tests/completeness:test_pg_transaction_boundary target passes all six cases against fresh PostgreSQL 16.15 (965187/0, independent 98c573/0). The sixth now rejects with M38061 after the fulfilled callback catches SQLSTATE 22012; observer/writer rollback and same-writer clean reuse also pass. The other cases retain commit visibility, callback rollback, caller savepoints, caller rollback and deferred-constraint COMMIT failure/recovery. The test does not expose the raw native COMMIT command tag. Each accepted run preserves all 6,384 product records and source/tool guards; the real-PG fixture’s observed registered processes and endpoints close.

Tracked Gnomon evidence, with public archive availability pending:

  • gnomon/tools/roma/evidence/pg-commit-outcome-db-first-failure-20261017/README.md retains the original 14-pass/three-failure package attempt.
  • gnomon/tools/roma/evidence/pg-commit-outcome-db-20261018/README.md retains the 17-case package pass.
  • gnomon/tools/roma/evidence/pg-commit-outcome-20261018/README.md retains the six-case sentinel-SQL pass.

These results establish the tested official-adapter/API SQL boundaries, not full PostgreSQL Graph module/service behavior or arbitrary custom-query-client conformance. No blanket rule for all caught errors or pending operations is selected. Full PostgreSQL module-resource persistence, registry/audit/activation, concurrency, proof/formal release and published integration remain open. Separately tested Site c029b86d6a77cbe94891a83991916fefa396ed9c consumes Spec dd6620275d5f6e4a797535cc31732a27e9c2db2f: corrected build 6d87e0/0, independent 064078/0, emits 395 HTML pages including 209 evidence pages and passes the existing route checks. The initial a0dc86/1 five-link failure is preserved before the finite source-link repair. This source-status refresh is not another Site build, browser acceptance or publication.

The real-installer persistence regression at Gnomon 713420af34abb299ef7105ec81bce294356ff527 passed the ordinary Roma //graph/tests/graph-module-engine:test_node target: 29 Vitest files and 231 tests, including both new filesystem/object-store cases. It uses the real engine, registry, object installer and default template executor. A fresh observer reads persisted core.module, core.module_manifest and resource-type rows: the first two successful modules remain visible before the third module’s real template-call failure; an object written inside the failing module is absent after rollback. With continueOnError=false, the fourth module is not installed; with true, it commits successfully. The archived actual evidence (Gnomon source: gnomon/tools/roma/evidence/rfc0041-filesystem-persistence-20261017/README.md; public archive link pending) records Root 3da076/session28494 → 6d3fef/0, Roma child 81366, independent acceptance a55ce9/0, and the unchanged 6,372-file product snapshot. This establishes the tested standalone module commit and rollback behavior on the no-savepoint filesystem backend. It does not establish the same module-resource behavior on PostgreSQL or inside caller-owned transactions.

The same ordinary Roma target at Gnomon 9d2b93386890a98d667c92b903d30bad50ee3f54 passed 29 Vitest files and 233 tests, including four filesystem cases in 199 ms (full Vitest duration 4.39 seconds). The two added cases cover both continueOnError choices: inside a caller-owned filesystem transaction, the caller catches the no-savepoint capability-check rejection, retains its earlier write, performs a later write, and commits both. Fresh observers verify both caller writes are durable, no module, manifest or resource-type artifact was created, and no template or content input was invoked. The caller-refusal evidence (Gnomon source: gnomon/tools/roma/evidence/rfc0041-filesystem-caller-rejection-20261017/README.md; public archive link pending) records Root 6de9c8/session10076 → c0d20d/0, Roma child 91202, independent acceptance 39ec98/0, and 6,373 unchanged product rows. This closes refusal before module writes on this no-savepoint backend. Full installer behavior inside a supported-savepoint caller transaction remains open.

The required-dependency planning gate passed at Gnomon 73486f7226b487d317fc5a6c39a541de54b41fdb: the ordinary Roma engine target passed 29 files and 239 tests, including six real-resolver cases. M41002 fails missing-required-dependency planning before module work for both continueOnError choices, retains complete failed/skipped reporting and order, and reports the missing module, dependency and requested version. Existing installed, in-batch and optional-dependency controls remain valid. The required-dependency evidence (Gnomon source: gnomon/tools/roma/evidence/rfc0041-missing-dependency-planning-20261017/README.md; public archive link pending) records independent acceptance eb1e66/0 and 6,373 unchanged product rows.

The subsequent ordinary Roma engine target at Gnomon 87b77a0d41341b03db322e06f7559940806e02c9 passed 29 files and 245 tests. Six added real-resolver/service cases check M41001 cycles, M41004 exact duplicate (namespace, version) pairs, and a valid dependency-first/lexical-order control. Failures occur before install or reflection-refresh dispatch; cycle diagnostics retain the same error instance/class and deterministic cycle. The planning-diagnostic evidence (Gnomon source: gnomon/tools/roma/evidence/rfc0041-planning-diagnostics-20261017/README.md; public archive link pending) records independent acceptance 511d02/0, 6,374 unchanged product rows and true source/tool guards. These are unit and prewrite planning checks; they do not settle different versions under one namespace, optional-dependency M41003, full SemVer selection or deployment policy.

That engine run consumes the corrected callback-lifetime client. Its separate ordinary Roma //graph/tests/graph-client:test_node run at Gnomon 355de169faa945592161ecf684082356872a2083 passed eight files and 42 tests, including fifteen M38060 lifetime cases, with 6,374 unchanged product rows and true source/tool guards. Both transaction entry points permanently close each callback’s client and retained capabilities when the callback settles; later callbacks cannot revive them, child closure preserves the outer callback, and materialized data remains usable. See the client evidence (Gnomon source: gnomon/tools/roma/evidence/graph-client-callback-lifetime-20261017/README.md; public archive link pending) and its preserved first compile-failure attempt. This establishes scoped client lifetimes and this engine consumer run, not M38061 error-propagation semantics or all custom-facade compilation.

The runtime normalized-manifest hash/canonical persistence subcheck passed on the real filesystem installer at Gnomon 9a323f63c84573b4bf9d9904d7da6f6ee30128be. Root 376198/session33516 → 514bbe/0 ran the ordinary engine target: 29 files and 246 tests, including five filesystem cases in 465 ms and all seven transaction cases in 18 ms (full Vitest duration 3.47 seconds). Independent acceptance 6dbe85/session57971 → 47a1cb/0 joins all 35 artifact bindings, 6,374 unchanged product rows and true source/tool guards. The hash-persistence evidence (Gnomon source: gnomon/tools/roma/evidence/rfc0041-module-hash-persistence-20261017/README.md; public archive link pending) retains the first false target at 14a84d6850e4ed77bcf21310b36f0201d7eb884a (c986f3/session73788 → 11077a/1): its new filesystem case passed, but seven transaction cases failed because their required crypto-service fixtures were empty objects. The retry changes only those fixtures to the existing real NodeCryptoService; production hashing was not bypassed or weakened.

The actual hash case reads persisted core.module through fresh observers: annotations-only changes preserve the SHA-256, object or required-dependency changes alter it, and a fixed independent JSON/hex oracle checks ordinal numeric and Unicode key ordering. The existing module ID survives upserts; stored canonical changes from MAJOR 3 to 4 while full SemVer version remains. Reinstalls explicitly use force: true. This closes tested Node runtime canonical/linkHash storage on the filesystem backend only. The later approved source below implements compile-time hash production and its complete artifact handoff; remaining consumer/wire integration acceptance remains pending. Automatic hash-triggered reinstall, its same-canonical no-op interaction, registry authority, PostgreSQL module-resource persistence and full framework conformance remain open.

The owner-approved complete compiled-artifact source is adopted at Gnomon 0ac0abad2ab29c4d97a6d04c100163b2f3f20924, with its authored contract at Spec a8183f65ef104ac19aa14d5abdb804054d41b6cf. The required per-module envelope carries the full normalized resolved manifest, exact derived templates and manifestHash. It verifies captured input before installation activity, stores the complete artifact with the matching core.module.linkHash, preserves the installed pair on a skip, and carries the hash through output-only RPC tag 6. The current routine integration repair is adopted at Gnomon 2dfe35d9d0d2a5c0ca0943169c47fddea83bed74: two metadata-only dependency resolvers regain their existing authored-manifest types, template lookup uses the transaction’s exact kind/key primitive, and its existing failure-test stub matches that primitive. The compiled-only install boundary is unchanged.

The owning auth BUILD generator and pinned pnpm 10.28.0 offline lockfile-only workflow passed: four generator fixtures, BUILD write/check, and the exact compiler workspace link, adopted at 4b355e289. This is generated dependency metadata, not compiled-artifact behavior. The descriptor-generation attempt (Gnomon source: gnomon/tools/roma/evidence/rfc0041-compiled-bundle-proto-generation-first-failure-20261017/README.md; public archive link pending) failed at an Abseil generated module-map input before the adapter/updater ran; neither descriptor output changed. The first model attempt (Gnomon source: gnomon/tools/roma/evidence/rfc0041-compiled-bundle-model-first-failure-20261017/README.md; public archive link pending) failed on the two engine TypeScript seams before any target or test body ran (zero observed cases); independent review retained all 40 artifact bindings and 6,382 unchanged product rows. The three production repairs and one test adapter above address those exact diagnostics. The model retry at that exact source is independently accepted: Root ea9c7c/session42188 → c74d92/0 passes 20 files, 318 tests plus one unchanged pre-existing skip (319 total), including all seven compiled-artifact cases in 20 ms (Vitest duration 1.62 s). Independent 4897ae/0 joins all 41 artifact bindings and 6,382 identical before/after/current product rows, with true source/tool guards. This accepts the selected model target, without granting aggregate framework acceptance.

The first compiler run (Gnomon source: gnomon/tools/roma/evidence/rfc0041-compiled-bundle-compiler-first-failure-20261017/README.md; public archive link pending) at 2dfe35d9d failed with 25 files, 193 passing and two failing cases out of 195; independent fdcfc2/0 retains its actual result. Both failures were fixture adaptations: the constants-options case had lost its unrelated nonempty template when the invalid old envelope was replaced, and the new permission-projection case used a non-core namespace prohibited by the existing schema. Commit 3c13fdf9 restores a valid derived inline template while retaining the include-toggle assertions, and uses exact core ownership for that permission fixture. No production validation or generator behavior changed. The retry at 5b223cb2 is independently accepted (848040/0, 0197bb/0): 25 files / 195 cases passed, including all seven handoff cases in 39 ms and both constants-options cases in 25 ms (Vitest 4.02 s). All 58 artifact bindings and 6,382 before/after/current product rows join with true guards. This selected target excludes integration/e2e, compiler.test.ts, artifacts.contract.test.ts, node-coverage.test.ts, stdlib-bootstrap.test.ts, core-ui-resource-policy-coverage.test.ts, build.command.test.ts and plan-apply.command.test.ts; package-local compiler tests remain separate.

The first engine run (Gnomon source: gnomon/tools/roma/evidence/rfc0041-compiled-bundle-engine-first-failure-20261017/README.md; public archive link pending) at 5b223cb2 failed with 29 files, 255 passing and three failing cases out of 258 (e0d2fa/1, independent 1b3681/0). Its bridge mocks lacked getAllBundles, and its direct template fixture lacked the captured envelope and transaction while retaining the old singleton-store expectation. Test-only 3af6949b corrects those fixture APIs. The second engine run (Gnomon source: gnomon/tools/roma/evidence/rfc0041-compiled-bundle-engine-second-failure-20261017/README.md; public archive link pending) at 3af6949b passes 257 cases and fails one before template dispatch (0fe24a/1, independent 5c58e5/0): strict finalization rejects the fixture’s missing visibility: 'public' default. Its new transaction assertions are not reached. Commit d1e27eb6 supplies that exact existing schema default without weakening finalization or changing production; both failures remain preserved at 3af6949b and 14da947d.

The engine v3 result (Gnomon source: gnomon/tools/roma/evidence/rfc0041-compiled-bundle-engine-20261017/README.md; public archive link pending) at exact source 14da947dc94f74dc8a24698db64829d12f7ee1d9 is independently accepted: Root c8938d/session45092 → 310d9f/0, Roma 92244, 29 files / all 258 cases passed, Vitest 2.99 s. All eight real-filesystem persistence cases pass in 423 ms, thirteen transaction-body cases in 41 ms, ten bridge cases in 22 ms and three template-deployment cases in 10 ms. Independent 47bc44/0 joins all 48 bindings and 6,383 unchanged before/after/current products with true guards. Production is unchanged since 2dfe35d9d; the additional product is the adopted authored documentation guide. This accepts the selected local engine target, including the captured-template transaction assertions and filesystem persistence; PostgreSQL module/service acceptance, registry/audit and broader integration remain open. The remaining consumer results retain separate scopes. The package-local compiler target at d27d5f9d passes six files and all 122 cases (1be647/0, independent c7dcc8/0), without covering the other compiler target’s exclusions or implying a test-types target in its hand-managed BUILD. The modules tRPC target at that source passes seven files and all 23 cases (68217a/0, independent 3a6c02/0): real finalized bundle forwarding uses a mocked engine. Its three installModules cases take 7 ms (Vitest 601 ms), with 46 bindings and 6,383 unchanged products. The Auth retry at a1bceb6f4 passes two files and all five cases (166f24/0, independent b3e908/0): four resetModuleGraph cases in 25 ms and router smoke in 667 ms (total 821 ms). Its first runfile failure remains preserved; the BUILD-only repair reuses the existing root store. Uploaded-source materialization and strict handoff retain dependency doubles, not authenticated public RPC/service acceptance. All 6,383 products and guards remain unchanged. Model, selected compiler, engine, local compiler, modules and Auth therefore have six distinct scoped passing suites; counts are not combined into broader acceptance.

The server target fails before any observed test body or named target result at d27d5f9d (96f614/1, independent 16df18/0, 55 bindings). The provider target at a1bceb6f4 likewise fails before cases (56a5c1/1, independent 4f01dd/0, 48 bindings). Each retains 6,383 unchanged products with true guards. Their 631-byte raw and 284-byte gzip logs are byte-identical to the preserved full descriptor-writer failure: Abseil CppCompile is missing generated log_severity.cppmap. These observations classify build failures, not failing server/provider cases. App-boot remains unexecuted pending the shared proto producer repair; its direct graph-proto-ts dependency reaches the same resolved-protoc rule, but no boot pass/failure is claimed. Provider/app-boot fixture type checks are also unrun. Remaining server/provider, boot/wire, PostgreSQL full module-service and full framework gates stay open.

Spec 28787b8bcca5430588c2db094acf776f603f0aab adopts the two generated RFC40/41 references and scoped review-fingerprint refresh. Ten newly mapped envelope requirements remain unproved, with no added logical axiom or borrowed proof. The data-only emission and query checks do not establish proof or runtime conformance. The current ledger retains 6,962 reviewed nodes, 2,259 mismatch verdicts and zero unreviewed nodes; final clean-HEAD generation/preview remain required, and formal release is held. The selected compiler and filesystem-engine checks do not replace the remaining provider, boot and RPC consumer acceptance gates. The linked-set linkHash is a different domain; hash consistency does not authenticate compiler origin.

Before that adoption, source inspection found reduced fresh-compile, persisted and RPC projections: moduleManifestEnvelope retained only namespace/version/objects and synthesized metadata, while installed serving retrieved latest templates separately. The approved source replaces those lossy paths with the captured full artifact; it does not establish their runtime correctness. Qualified-to-bare template lookup and verified candidate-seed fallback retain their existing gaps, so installed dependency-closure authority is not claimed. Observed-pair read validation also does not establish snapshot isolation or concurrent-read availability. Automatic hash-triggered reinstall and its same-canonical no-op priority remain separate decisions.

The exact-artifact/same-canonical upgrade counterexample is reproduced at tests-only Gnomon efcb2aafbc4c8b2fb1fbb19eb75c1c3a9cd89536: the existing engine target runs 260 cases, with 258 passing and both appended cases failing (a0e88c/1, independent 2374dd/0; 6,384 product records unchanged). The real filesystem case changes fresh-observer rows for the identical installed bundle at line 531; its following no-action assertion is unreached. The injected stale-hash case fulfills instead of rejecting at line 1072; later refusal assertions are unreached. A narrow source repair is adopted at Gnomon 652cd4b2642a0c8ca7fe0f623cb41ba152e4b885 (lifecycle-service SHA-256 965febaa26cf75ea39958a0272b758c6aaca5b51313707cb64b30f07e8f6265a). It verifies the captured target and installed artifact before an equal-hash early return. The existing engine target now passes all 29 files and 260 cases at that repaired source in 3.22 seconds (a24fe4/session71587 → f1c749/0, independent d3113e/0; all 49 artifact roles and 6,384 unchanged product records). The filesystem file passes nine cases in 528 ms and the lifecycle file passes 23 in 16 ms. The fresh-row equality failure is resolved; the previously unreached no-action and hook/install/event refusal assertions now pass. This closes only verified identical-hash/same-canonical execution and strict injected stale-hash rejection. The first failure and accepted retry are retained at these local source paths, with public archive availability pending:

  • gnomon/tools/roma/evidence/rfc0041-equal-hash-lifecycle-first-failure-20261018/README.md
  • gnomon/tools/roma/evidence/rfc0041-equal-hash-lifecycle-20261018/README.md

Changed-hash priority, automatic reinstall, force/uninstall, no-op provenance, installed-state authority, full PostgreSQL module behavior and concurrency remain open.

The §8 phase-order regressions failed against unchanged production at test-only Gnomon ee7f02756dd812e8ee4ba018035e7a27f40b1afa: Root d687d7/session62186 → f29207/1, native child 27235, ran 29 files and 249 cases (245 passed; four ordering failures). Independent read 37ac20/0 joins all 37 artifacts and 6,373 unchanged current product records. The failures reach the ordering assertions in install-module.transaction.test.ts:588/624; later absence assertions are unreached, so this run does not prove prevention of later-phase writes. The first-failure archive (Gnomon source: gnomon/tools/roma/evidence/rfc0041-object-phase-order-first-failure-20261017/README.md; public archive link pending) preserves the evidence. The owner subsequently approved the deep-exported bulk hook’s change from one global call to per-phase calls with context subsets. The normalized two-production-file repair is adopted and tested at Gnomon 9ac77c6060aa4fd06adba8a92ec71aa38bf6eefd: Root 395637/session17590 → 314dae/0, Roma child 42674, passes all 29 files and 249 tests in 3.08 seconds. Independent acceptance e8b4ea/0 joins all 37 artifact bindings and 6,374 unchanged before/after/current product records. The repaired phase evidence (Gnomon source: gnomon/tools/roma/evidence/rfc0041-object-phase-order-repair-20261017/README.md; public archive link pending) records that all four ordering regressions now pass, including the previously unreached later-phase absence assertions. The fixtures use real primitive and template handlers, with statement/policy dependency doubles. Operation and data-source handlers remain no-ops; full nine-kind storage, qualified-resolution authority and PostgreSQL module acceptance remain open. This phase-only result does not test the later compiled-artifact source adoption.

The tested diagnostics and prewrite planning gates, scoped client lifetimes, five PostgreSQL API/backend SQL cases and four no-savepoint filesystem module-persistence and caller-refusal cases are closed within their stated scope. Full engine conformance remains unverified. PostgreSQL module-resource/install-plan/registry persistence, install-level audit records, compile-time manifest hash handoff and automatic reinstall/lifecycle selection, complete object-kind registration and storage, executable linker/provider behavior, agent activation atomicity, concurrency and the published graph integration checks below remain open. The focused SQL test uses a sentinel table, not Graph module resources, a registry or an install audit writer; state-model and recording-client tests retain their own scope.

Historical audit — October 4, 2026

Status: extracted framework source audit, 2026-10-04. Owner: Module Platform, Graph Runtime, and Spec. This is a first-release gap record, not a claim that the authored RFC has been implemented.

At that audit, the graph-module-engine followed parts of the older module model. The RFC-0041 projection review and source inspection found these concrete differences:

RFC-0041 contractExtracted source at the October 4 audit
§6.2: each successful module commits before the next, and a later module failure cannot roll it back.installation-service.ts calls runBatchInTransaction, which opens one outer transaction for the batch; savepoints isolate handled module failures, but earlier modules do not commit until the whole batch callback finishes. The no-savepoint branch cannot isolate a failed module for continuation.
§7.3 and §11.1: hash the normalized manifest, exclude metadata.annotations, and store the hash and canonical version on core.module.install/hash.ts exposes hash helpers, but the module materialization template in install/install-module.ts writes only name and version. linker-service.ts hashes a linked module set, a different object, and does not supply the installed manifest hash.
§8: object failures emit M41020 with the cause.install/install-module.ts now throws a contextual plain Error; the explicit M41020 diagnostic is absent. The throw allows rollback, but the diagnostic contract remains open.
§8: operations register without executing; object order is predicates, resource types, templates, operations, resources, policies, calls, statements, data sources.object-installer-service.ts maps OPERATION to a no-op. INSTALLATION_TYPE_ORDER in install/install-module.ts begins with ID ranges, resource types, resources, and predicates; it has no explicit template-call slot and puts templates after statements.
§9: after installers, the linker binds providers to tokens, resolves template calls, defers resolvable references, and applies provider precedence with M41031/M41030.ModuleLinkerService.linkModules currently builds a deterministic linked-set artifact and hash. A source search found no call to it in the module engine or graph server, and its implementation contains none of those runtime bindings or diagnostics. Template calls are instead executed directly by the object installer.

The code changes in 279bbaf76 and 65d300610 removed PostgreSQL audit suppression and made object/materialization/hook errors propagate for rollback. The graph-module-engine Bazel target passed 28 Vitest files and 216 tests. Those tests do not cover the full contracts above. The separate audit and atomicity record defines the remaining failure and audit evidence.

Acceptance

  1. The no-savepoint filesystem branch passed the two real-engine cases above. Complete the PostgreSQL branch with two successful modules and a failing third module, checking actual module, manifest, object and registry rows. Verify commit boundaries and continueOnError by querying persisted graph state; the five current SQL cases use a sentinel table.
  2. The filesystem hash/canonical readback subcheck passed above. Verify the adopted compile-time producer and serialized/provider/persisted/RPC handoff with actual runtime results, including PostgreSQL values. Test automatic hash-triggered reinstall and same-canonical no-op/lifecycle selection; forced reinstalls do not prove them.
  3. Exercise every object kind in the authored order, including operation registration and qualified template calls. Verify M41020, M41030, and M41031 at their declared phases with structured diagnostics.
  4. Test provider override, category precedence, install order, ambiguity, deferral, and unresolved references against a real DI registry. Confirm the linker executes in the specified phase and does not merely emit a linked-set artifact.
  5. Build the published graph dependency closure and rerun the spec/RDF/site gates before making an implementation conformance claim.