RFC-0041 · Design Q&A
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.
Q: Does a same-major manifest hash change reinstall or no-op?
Status: Open
Source: RFC-0041 §§5.2–5.3, 7.3, 13 invariants 4 and 9 Issue: The lifecycle table makes every same-major target a no-op and skips object phases, while a changed manifest hash must trigger reinstall even at the same version. The policy also says no inputs other than version comparison influence selection. Recommendation: Add an explicit same-major update strategy based on content hash and version, with a defined migration and hook policy; reserve no-op for identical installed content.
Q: How is stable identity unique for two same-kind objects in one namespace?
Status: Open
Source: RFC-0041 §§7.1, 8, 11.2 Issue: The stated stable-id formula defaults scope to an object’s namespace path or local name. Two resources with the same namespace path can collide, while per-kind examples use the local name. The rule for scope selection and collision detection is absent. Recommendation: Include the local object name in every default stable id, define escaping and canonicalization, and reject duplicate derived identities before writes.
Q: When are symbolic references resolved relative to object installation?
Status: Open
Source: RFC-0041 §§3, 8.1, 9.1–9.2 Issue: Template calls and statements execute in phase 5 but the linker resolves their template, predicate, and resource references only in phase 6. Deferred links to later modules cannot be completed inside the already committed module transaction without another mutation phase. Recommendation: Resolve required declarations before dispatch, use explicit placeholders only where the backend supports them, and define a transactionally safe end-of-plan linking phase.
Q: Is forced reinstall atomic with removal of the previous installation?
Status: Open
Source: RFC-0041 §§3, 5.4, 6.1, 12
Issue: forceReinstall uninstalls existing graph state before a fresh install, but the transaction boundary starts after lifecycle selection and its relation to uninstall is not specified. If the new install fails, previous state may already be removed.
Recommendation: Put uninstall and reinstall in one transaction where feasible, or define a recoverable staged upgrade with rollback evidence and an explicit operator acknowledgment.
Q: How can a read-only preUninstall hook clean up derived state?
Status: Open
Source: RFC-0041 §§10.1, 12 Issue: The hook is described as read-only but also allowed to perform cleanup of derived state. These requirements cannot both hold if cleanup writes to the graph or an external system. Recommendation: Restrict the hook to validation and move cleanup into a transactional uninstall phase, or define a separately named write-capable hook with its failure and rollback rules.
Q: What does forced removal do to graph referential integrity?
Status: Open
Source: RFC-0041 §12; RFC-0036 §9.2
Issue: force: true accepts dangling statements from other modules, while the graph contract requires statement endpoints to reference existing non-deleted resources. The uninstall sequence also removes ownership statements before it uses them to identify owned resources.
Recommendation: Preserve the owned-resource set before mutating ownership links and either reject referenced-resource deletion, cascade affected statements with provenance, or define a tombstone model consistent with RFC-0036.
Q: How can module continuation preserve a caller’s activation transaction?
A: The approved module boundary owns a transaction when standalone and uses a savepoint when nested. A failed standalone module rolls back before continuation; a failed nested module rolls back to its savepoint without committing the caller. Nested execution without complete savepoint support rejects before module writes. Lifecycle failures escape both entry points; only safely rolled-back installer failures may become ordinary failed results. The caller retains the outer activation scope required by RFC-0005.
Source: RFC-0041 §§6.1–6.2, 13; RFC-0038 §3.4; RFC-0005 §4a; approved Gnomon implementation fed0662ece0d3ecd6774469085cef261d303ad9e
Issue: The old engine transaction covered the whole batch; a simple per-module call through the joining API could instead leave failed nested writes pending. The adopted source repair has ordinary Node regression evidence. The focused PostgreSQL 16.15 API/backend SQL target at Gnomon 091f1f0060fff18cf7168e23a77a88a0f5682beb separately passed one Vitest file and all five tests: owned commit visibility, callback rollback, caller savepoint isolation, caller rollback and deferred-COMMIT-failure recovery on the same writer/backend, with an independent observer. Actual scoped evidence (repository access required) binds Root 4328dc/session10051 → c9f827/0, Roma child 55899, and independent acceptance c51d53/0 plus dbe597/0. The real-engine filesystem no-savepoint cases at Gnomon 713420af34abb299ef7105ec81bce294356ff527 also passed within the 29-file/231-test ordinary engine target: fresh observers read the first two committed modules, manifests and resource-type rows before a third module’s failure, verify its staged object rolls back, and see a fourth module only with continuation enabled. Filesystem persistence evidence (repository access required) records independent acceptance a55ce9/0. At Gnomon 9d2b93386890a98d667c92b903d30bad50ee3f54, the same target passed 29 files and 233 tests, including four filesystem cases in 199 ms (full Vitest duration 4.39 seconds). The two added caller-refusal cases catch the no-savepoint capability-check rejection inside the caller transaction for both continuation choices; fresh observers verify the caller’s prior and later writes commit, no module artifacts exist, and no template or content input is invoked. Caller-refusal evidence (repository access required) binds Root 6de9c8/session10076 → c0d20d/0, Roma child 91202, independent acceptance 39ec98/0, and 6,373 unchanged product rows. Refusal is verified; full installer behavior inside a supported-savepoint caller transaction, PostgreSQL module-resource/install-plan/registry persistence, install-level audits, agent activation atomicity, concurrency and other backends remain open.