Library and sandbox qualification
Owned-game launching and optional sandboxing remain Partial. The source implements saved per-installation launch configuration, profile/save transitions, exact-install store hooks, durable descendant observation and paired performance capture. Automated fixtures establish these contracts; real-game and distribution qualification is still required before treating modde as a qualified Lutris replacement. Real-game sandbox performance remains unmeasured.
The required implementation target is native Linux. Flatpak host-launch integration is a follow-up. Performance measurements are report-only, as selected by the user on 2026-10-01: retain measured deltas and evidence without a negligible-overhead pass/fail verdict.
Current qualification, 2026-10-06
Merged default source
PR #3 merged normally as
db5b249feba133587a9289b66ed6da5191d3517d, tree
4fdf1d1fc999c55794014fa1bc60b5bf8e942f58. Its
default CI passed all
five jobs and its Nix matrix
passed all four builds. Independent guarded selection matched every retained
derivation/output pair to this exact revision and verified all sixteen artifact
file hashes and the source NAR. Receipt:
/data/nvme0/can/ProjectState/generated/modde-ci-source-closeout-db5b249-20261006/default-independent-20261006T064801Z/receipt.json
(SHA-256 fe9dface81759c84e6b795714f55ee1c355b26f988dae509f4f8632daeb84d37).
The Pages run failed
before deployment while compiling Plinth. Harbor’s compiler wrapper reported
no managed cache transport is available and exited 75. The producer follow-up
selects the existing Plinth cache repair and requires a successful Modde site
build and deployment at its own candidate revision.
PR #2 closed as superseded after the operator selected preservation. The exact capability mapping retains nine default members, six minimal members, seven package inventories and all 25 named release steps, including optional Windows Authenticode signing. Optional Greptile review remained unavailable; native protection imposed no mandatory provider review. Release credential readiness and actual channel delivery remain separate gates.
Producer dependency follow-up
The follow-up pins Plinth 6f17df07bf3c3a33a753c459872e035774a1a27d beneath
visual-rubric/plinth. The root Plinth alias and Simit follow that same source;
Visual Rubric retains its existing 7e63e9458ff65327b5a1501fdc7723a07166702d
revision. The managed declaration/lock update passed all-output evaluation on
x86_64-linux, aarch64-linux and aarch64-darwin; check builds were explicitly
deferred. Plinth’s upstream CI and Pages passed, but Modde’s changed site still
requires its own package and deployment acceptance.
Cryoglyph’s published 0.1.0 release requires vulnerable lru 0.16. Upstream
fixed that dependency in a7a4b4b5a9fe9bf621c5c858489520fe0661ae5f, alongside a
graphics stack incompatible with Iced 0.14. The follow-up therefore retains the
published WGPU 27/Cosmic Text 0.15 API and backports only the fixed lru 0.18.2
declarations. scripts/vendor-cryoglyph.py --check verifies the source archive
checksum and every vendored file; upstream Rust/WGSL sources and licenses remain
byte-identical. The Nix production source includes this patch; native, cross and
test derivations compile real sources with cargoArtifacts = null, preserving
Harbor’s rejection of unsafe dummy dependency-only builds.
Native verification in an isolated SQLite home passed workspace all-feature,
all-target compilation and warnings-denied Clippy. All-feature tests passed
2,095 cases, with zero failures and two ignored; minimal UI tests passed 201
cases with zero failures. cargo audit --json passed with zero vulnerabilities,
including removal of RUSTSEC-2026-0253. Four unmaintained and two yanked-package
warnings remain recorded. These are local source/dependency results; changed
production derivations and the exact committed hosted candidate require their
own receipts.
Approved Canix runtime cohort
The approved producer revision is
0d3ce666e2362555b90d4d37b084cc8d86450eb9, published on
integration/native-library-gpu in PR #3.
Its producer CI
and native package CI
passed. Independent verification retained the hosted checkout/tree identities
and reports 2,092 passed tests, zero failures and two existing ignored tests:
/data/scratch/tmp/opencode/durability-modde-0d3-independent-producer-verification.json.
These results qualify that revision; later documentation or implementation
changes require their own checks before a replacement pin is selected.
The historical Canix consumer
8240764b480d0fff6aab77ef50f8767b5941676e realized and privately published
the actual Home Manager wrapper and native package:
- Wrapper:
/nix/store/zdyx23lxxyzbj4z2gnzy5h4h8amwfxjj-modde-desktop-runtime. - Native package:
/nix/store/8vbqbqbdp86qvrwc0yscicsz6ll2yfr2-modde-0.7.0. - Binding:
/data/scratch/tmp/opencode/modde-atlas-repaired-committed-runtime-binding-20261002.json.
The actual-wrapper fixture used isolated SQLite/configuration/data directories
and real bubblewrap. It passed successful completion, detached descendant
failure, signal handling, failed inner-exec recovery, private/redacted diagnostics
and journal cleanup. Durable raw statuses were 0, 1792, 15 and 256
respectively. The binding records the desktop entry, both wrapper defaults,
compiled producer revision, NAR/deriver identities, launch receipts, retained
output root and verified private publication. Defaults are PostgreSQL through
/run/postgresql, render node /dev/dri/by-path/pci-0000:03:00.0-render, log level
info and retention of 50 launches. The fixture’s SQLite override does not
establish live PostgreSQL or real-game save continuity.
On 2026-10-04, read-only reverification confirmed those retained artifacts and
receipts:
/data/scratch/tmp/opencode/modde-native-closeout-historical-verification-20261004T194904215074Z/receipt.json
(SHA-256 4c272affec938bba414c72857199992683aa4a0a677a285ee664d4c0134ea81d).
This remains historical component evidence. A successor Canix source requires
measured package, wrapper, defaults and dependency-context equality before
execution evidence transfers. Equal source files alone do not establish it.
Canix qualification runs locally on Atlas; its hosted Actions are
skipped-by-operator. The coordinator selected
3b29bb33d125381f9d34b4f18657755cbe4a0379 for scoped non-activating qualification
after reviewing the complete successor delta. The predecessor online VM failed
because its stdout-only capture missed the intentional stderr stop; the corrected
fixture preserves all 13 lifecycle assertions. That failure and earlier failed
or cancelled attempts remain retained. Selection authorizes component checks,
not production adoption or activation. Final composition, full local registry
qualification and the production writer window remain coordinator-owned.
Live game testing starts after those recovery/adoption/activation gates. At runtime entry, reconfirm Modde inactivity and the actual foreground marker/pause policy. Native direct, Steam, Heroic, Proton/UMU and manager paths need observed rendering, interruption recovery and save-continuity receipts. Eight alternating off/on measurement pairs require about 80 minutes of gameplay plus setup; retain paired deltas and variability without an overhead acceptance verdict. PR review/merge readiness remains separate from successful component CI.
Selected Canix binding and component acceptance
The selected 3b29bb33… combined binding and coordinator post-checks passed.
Independent verification covers source NARs, ten standalone/embedded homes,
41 recovery-record declarations, separate configured and final-service PostgreSQL
packages, opaque derivation graphs, verified private canix-fleet publication
and physical/registered managed roots. The retained inventory contains 72 exact
derivations: 70 explicit selections plus the measured manager package/config
output contexts. Evidence:
- Sealed worker handoff:
/data/scratch/tmp/opencode/durability-3b29-green-binding-worker-handoff-20261004T220520080406Z.json(SHA-25627fbdd313cb10fe9edaa55b5b276d6dc65a49a92e04b8bf53c560e165d9bdfa6). - Combined binding:
/data/scratch/tmp/opencode/durability-3b29-coordinator-combined-context-v2-20261004T213013602808Z/binding.json(SHA-25681e06c9737a36177488b9b2254231933c07ac1c5f10ff07b7ccfe68aa53fa0b6). - Independent binding/root verification:
/data/scratch/tmp/opencode/durability-3b29-bound-selection-context-v5-independent-20261004T220352406667Z/receipt.json(SHA-256f0972b6f996b28b023975277f63b3b64985aed2ad1cf84d6700bb5d56b0ff49f). - Actual rendered production shell: unsuppressed ShellCheck and all twelve cleanup
cases passed in
/data/scratch/tmp/opencode/durability-3b29-authoritative-full-script-native-20261004T215948931439Z/receipt.json(SHA-256f6bab51b5ac28f6b95f8bf3f9b01a11d204f91a99a3b353c05b8cb71e9172933).
Measured equality transferred the historical Modde wrapper/package and remote-policy
results to this selection. The coordinator independently accepted artifact/NAR,
deriver/reference, five-input cohort, defaults, database ownership, four launch
statuses, remote opaque graph, private publication and root equality:
/data/scratch/tmp/opencode/durability-3b29-modde-remote-transfer-independent-20261004T221931871448Z/receipt.json
(SHA-256 940b25629498d6d1220392bbb945bceb654f312dba4b24b19a382b6eefbd99f0).
The launch evidence retains its disposable SQLite/real-bubblewrap scope.
Modde’s storage cohort uses final-service PostgreSQL
/nix/store/rnyi9ykl60rghq1z2nbvl29jljjchak7-postgresql-and-plugins-18.6.
The separately verified configured PostgreSQL package has a different output
(3hc…); it must not substitute for the final-service binding.
Corrected online VM result
The sole actual corrected pg-online-backup-only retry passed on 2026-10-04.
The reviewed v2 runner realized the selected opaque graph with normal capacity
admission, one job/two cores and no further flake evaluation:
- Derivation:
/nix/store/cn472qzq1g4ng8d9fm9pskpccll4jgap-vm-test-run-canix-pg-online-backup-only.drv. - Output:
/nix/store/s219r8labpnic76scwfz4y9w6p8a53gk-vm-test-run-canix-pg-online-backup-only. - Actual terminal:
/data/scratch/tmp/opencode/modde-atlas-3b29-bound-pg-online-backup-only-immutable-executor/terminal.json(SHA-256ff27fce292c58b09e49605533bae733f508b09e2fa332050ac1bc90fe88ff05f). - Worker post-terminal verification:
/data/scratch/tmp/opencode/modde-3b29-online-terminal-verification-20261004T223136116229Z/receipt-v2.json(SHA-256d08e7ea2ff6a53f3ec4a240bd400476286a7054dff0f249161e545eb3529333f).
JUnit reports one test, zero failures/errors/skips. The successful complete
rendered script retains all 13 lifecycle assertions, including intentional
pre-adoption stops, INT/TERM interruption cleanup, receiver restoration, retained
partial/acknowledged-row preservation, LAST_SUCCESS and four receipt hashes.
The complete 1,207-line VM log and exact private publication/physical managed root
were retained and rechecked. Captured intentional-stop stderr is checked by the
script’s assertions; the driver does not echo that captured text into its log.
The initial post-terminal log-substring assumption failure is preserved alongside
the corrected verification, as are the predecessor VM failure and admission-only
nonexecution. This is VM component acceptance, not production record acceptance,
adoption, activation or actual-game continuity.
Changed-hook Nomad VM and independent acceptance
The user explicitly reassigned the sole changed-hook Nomad VM to the coordinator. That exact selected graph also passed through one opaque managed realization, one job/two cores, no source evaluation and no capacity override:
- Derivation:
/nix/store/5s94bikx9yickgh6vrk0n28ymlwc233v-vm-test-run-canix-managed-pg-nomad-recovery.drv. - Output:
/nix/store/5k26h6clpais856ajachcprgwl3jk00j-vm-test-run-canix-managed-pg-nomad-recovery. - Actual terminal:
/data/scratch/tmp/opencode/durability-3b29-coordinator-pg-recovery-nomad-only-executor/terminal.json(SHA-2562e9c96bd085eee707be88a9df8b7ee3234357df5783891fad41cd04c7aad2db2). - Independent acceptance of both online and Nomad VM gates:
/data/scratch/tmp/opencode/durability-3b29-vm-terminal-independent-v3-20261004T224505764092Z/receipt.json(SHA-2569d7c8a498c7109dacf0f49399d15c298d435d8d50874fa9ae00ab1691b4bd823).
Both gates report JUnit one test, zero failures/errors/skips and exact private
canix-fleet publication/physical roots. Nomad’s actual bound production hook
checks export v2: exactly 2,000 files and 35 directories, all sizes/hashes and no
extras. The unchanged 41 production SQL checks plus an additive saved-reviews
fixture produce 42 equal records across source, local and off-host restores.
Backup, epoch, system, metadata, manifest, snapshot, WAL and contract identities
are bound. Cancel, wrong-export, stale-live-check and corrupt-metadata rejection
paths reach successful whole-script completion. Verification-adapter failures
remain retained: verbose live/ANSI logs differ from JUnit’s representation, and
the snapshot has no self-hash; its actual bytes/export and both restore hashes
provide the binding. No fixture/assertion changes or second VM attempt were used.
Recovered six-stage literal provenance
The exact 3b29bb33… top-level derivation closure selects all six Raven ETL unit
derivations. Read-only inspection recovered each unit’s literal rendered text
from structuredAttrs.text, establishing User=pink_raven, literal
PGUSER=pink_raven and source revision
a96a22ab6f554631e72068acdf22baab4bdbe622. These fields are byte-equal to
historical 224…, and every command context is represented by the selected
unit derivation inputs. Evidence:
/data/scratch/tmp/opencode/durability-3b29-six-stage-literal-provenance-20261004T225802091632Z/receipt.json
(SHA-256 d73f0b6fa9a96e91a62345c2fb5af55f97cd0bacb05129b89294e6482d59ecc4).
This closes the literal environment-projection evidence gap for 3b29… without
another flake evaluation or realization. Complete stage commands and contexts
still differ: selected Canix lease-launcher output/derivation wq42…/dsr3…
replace historical 17z…/b2i9…. Literal-field equality does not transfer
complete stage execution; the changed launcher and final composition still need
their own qualification.
Remaining closeout and production gates
Scoped local qualification is green. Physical production Nomad restore, the last
acknowledged-row boundary, writer window and final cohort remain unadmitted.
Production ETL writer authority remains unassigned/unconfirmed; VM reassignment
does not schedule or fence production. Final committed composition, packaged CLI
and full runtime/stage qualification, writer window and activation remain pending.
In particular, selected Raven stage commands use different Canix lease-launcher
artifacts from the historical accepted stage. Historical component acceptance
does not establish full-stage context identity across that change.
The recovered literal fields above are accepted for 3b29…; the final source
must bind its own stage fields, commands, CLI artifacts and contexts.
The coordinator must select an exact committed 40-character final Canix revision
and enumerate its production recovery contract. The accepted 41 production SQL
checks and additive VM fixture are specific to 3b29…. Observed Forgejo source
edits do not qualify a possible 49-record final contract. Each added or changed
record needs an explicit owner, source-backed assertion and qualification of the
changed scope.
Evidence transfers artifact by artifact. Record one of three dispositions:
| Disposition | Required evidence |
|---|---|
| Identical and transferable | Measured derivation/output, NAR, relevant inputs/defaults and execution-context equality, with retained publication/root evidence. |
| Changed and requiring qualification | Explicit delta and admitted qualification of the changed package, launcher, hook, recovery contract or runtime behavior. |
| Pending | The exact missing evidence and its owner; no acceptance inferred from source similarity or a previous pass. |
A changed recovery hook or SQL contract cannot inherit the prior scope’s VM
acceptance. Any required final-source attempt is separately coordinator-admitted;
the sole corrected online and changed-hook Nomad attempts for 3b29… remain
closed green. Qualification uses normal bounded admission and retains failed,
cancelled and intermediate evidence.
Production admission must bind the exact final artifacts, confirmed writer owners and protected workloads, UTC start/end/deadline, last-acknowledged boundary, acceptance criteria and the failure/deadline recovery action. Physical restore acceptance precedes adoption/activation; explicit writer release follows required post-activation verification. Modde gameplay and measurements require their own explicit runtime admission and cannot inherit an expired recovery window.
The coordinator’s 2026-10-05 storage recheck leaves an Atlas Btrfs blocker
unresolved. The operator’s fresh read-only scrub reproduced 60 uncorrectable
errors after scanning 4.97 TiB; corruption counters increased from 300 to 360.
The operator deleted the identified damaged Regicide test executable and
regenerated it successfully. Its complete read/hash passed, but no post-deletion
zero-error scrub or complete remaining affected-file inventory establishes
filesystem clearance. Historical evidence:
/data/scratch/tmp/opencode/durability-atlas-btrfs-persistence-20261005T053121748122Z/receipt.json
(SHA-256 6b27f32a8eb2b7f33a02dac173dba02c7d2b0062098f21d4dc14de7863f6aab8).
The coordinator/storage owner must provide accepted fresh storage verification
before production admission; successful historical VM gates do not clear it.
The initial PR #2 comparison used head
2cadcd139368620184fffd587df961684f90c60b and found nine per-crate CI workflows
and a release workflow absent from the then-approved PR #3 source. Its historical
non-supersession disposition remains retained:
/data/scratch/tmp/opencode/durability-modde-pr2-closeout-independent-20261004T222654796699Z/receipt.json
(SHA-256 50c8d44f324a3b661cee9e156a12bf7dedabf25b3ec105e29654a82d458bcfa5).
On 2026-10-05 the operator selected preservation. The qualified replacement
retained those capabilities and PR #2 closed on 2026-10-06, as recorded above.
GitHub CI/release source qualification, 2026-10-05
The managed Simit-only update 20261005T221507.725Z-2272603 promoted exactly
one lock node to 1f5a180b7bd8d3cf4f1e8c3a105b33c251fac0e1. All supported
systems passed output evaluation; native Nix check builds were explicitly deferred.
The Canix updater ordering repair b1561cf76b1aaa9a460180413444e1b887a099c5
selects the lightweight check inventory before full output validation against the
same immutable source. Seven updater tests, strict Clippy, an admin native build
and this actual transaction passed with the existing admission thresholds intact.
The patched native Simit generator regenerated CI, Nix-build and Pages workflows,
plus the separate release.yml. Both CI and release --check --diff gates passed;
actionlint with the real ShellCheck executable passed for all four workflows.
.github/scripts/check-crate-compatibility.sh preserves the nine isolated default
member checks, six minimal member checks and seven package inventories from
PR #2. Its 42 expanded native commands exactly match the accepted isolated R8
run, and all bound Rust/snapshot bytes remain unchanged. The additional
modde-xtask check is retained as a hosted gate because its pinned implementation
invokes cargo fmt; it has not been executed by the local agent.
All 25 named release steps from PR #2 are mapped: 24 are exactly equivalent,
while Windows packaging now retains the upstream conditional Authenticode
sign/verify sequence. Windows archives, individual executable assets and the
unsigned fallback remain available. GitHub releases, checksums/attestations,
Attic, APT, AUR, COPR, Flathub, winget, Homebrew, Scoop, Chocolatey and
announcements retain their source capabilities. Build-bearing Nix jobs are
limited to one job and two cores; the installable matrix has concurrency one
and includes the actual simit-ci-generator package.
Evidence under
/data/scratch/tmp/opencode/modde-ci-reviewed-20261005-coordinator/:
managed-simit-promotion-verified-r6.json: transaction, one-node lock delta and preserved foreign files;regeneration-managed-r7/regeneration.json: generator/configuration/workflow fingerprints and passed generated-file drift;workflow-lint-generated-r1/receipt.json: actual actionlint/ShellCheck result;pr2-capability-map-generated-r1.json: exact-head member and release mapping.
These source/native-generator results are followed by the packaged-generator,
exact hosted-revision binding and merge receipts recorded above. Release
credential readiness is unestablished: the
2026-10-05 repository secret and variable listings were both empty. The release
capability mapping establishes source preservation, while actual channel delivery
requires its own credentials and acceptance. Normal native forge integration
completed under the governing optional-review policy. The Greptile credit-limit
notice supplies no substantive review. Later documentation and implementation
changes need their own qualification and do not replace approved runtime producer
0d3ce666… or inherit earlier hosted results. Generated docs, active/history roots,
runners, queries and failed receipts remain retained for their scoped gates.
The dated sections below retain earlier implementation and failure evidence.
Repeatable gates
From the development shell:
cargo xtask library-qualify --jobs 4
cargo xtask library-qualify --containment --jobs 4
The first command checks the minimal core/games/sources/CLI builds, each of the nine individual CLI game features, save-recovery and installation boundaries, supervision/completion receipts, performance sample gates, bisect provenance and Library UI race guards. The second also runs real bubblewrap probes. It requires Linux, bubblewrap on PATH and usable user namespaces; unavailable containment fails the gate rather than skipping it. The development shell supplies bubblewrap and SQLite. SQLite is also required by manager integration tests.
cargo xtask check, used by the generated project-check CI step, runs these
qualification gates after the workspace checks and requires real containment on
Linux. checks.library-regressions supplies deterministic regression coverage
inside the Nix build sandbox; real namespace creation is tested on the CI host.
Validate that Nix gate with:
canix cache binary build .#checks.x86_64-linux.library-regressions --include-tests --no-push --max-jobs 1 --cores 2
Evidence from 2026-10-01
- Workspace/all-target compilation and strict all-feature Clippy passed.
- After GPU integration, the all-feature workspace suite passed 2,076 tests; the earlier default-feature suite passed 2,064 tests. Each run had two existing ignored tests. Fleetix’s all-feature suite and strict Clippy also passed.
- Minimal builds and the nine individual game-feature builds passed.
- The production native Nix package built successfully. Its installed CLI passed a sandbox lifecycle fixture: detached descendants completed before the session ended, and exit codes 0 and 7 produced durable aggregate evidence of 0 and 1792 respectively. This fixture does not establish rendering or game performance.
- Rust 1.93.0 passed workspace/all-target/all-feature compilation. Its invocation
removed the development shell’s nightly-only compiler flags; it did not use
RUSTC_BOOTSTRAPor bypass the declared MSRV. - Real bubblewrap 0.12.0 probes passed for read-only/writable grants, save and prefix writes, private HOME, inherited credential/session-bus exclusion, PID isolation and networking both enabled and disabled.
- A later Atlas desktop probe confirmed read/write access to the RX 7900 XTX
render node and its Wayland socket. The
qualify_gpuexample exercised stable PCI selection through modde’s direct and bubblewrap boundaries; both reported accelerated RX 7900 XTX OpenGL rendering, and Vulkan enumerated the RX 7900 XTX first with matching PCI IDs on both paths. Evidence is retained at/data/scratch/tmp/opencode/modde-gpu-rendering-20261001.json. This supersedes the earlier no-device observation and establishes a graphics diagnostic, not GUI-to-game or performance qualification. - Toolbelt render-route assertions passed, covering topology defaults, application
overrides, disabled defaults, invalid nodes and older-module compatibility.
A source-level evaluation with modde’s actual runtime Home Manager module
verified
MODDE_GPU_RENDER_NODEgeneration and override precedence. Fleetix’s example Pkl topology exported the new render-route contract successfully. - The scoped Fleetix GPU revision
c79c9a3902f746ec17f56de37f2c351e2dcdfbc3passed hosted CI, including flake checks, both feature suites, MSRV, audit, documentation, strict Clippy and packaging. Toolbelt revisionafd6a7ced2ebea3b98705ca875d26a7acd2e5792passed the exact pinned GPU/backend/ browser Nix evaluations and cross-adapter override assertions. Its broader CI failed because published Toolbelt trunk already expects Fleetix’s missingpublicationAddressIntents; the same failure was reproduced on unmodified trunk2faa0dd33f17386fb9123c2015bf6593d0159e40. The separate browser smoke job also failed withbinary is not a Firefox executable; its derivation is byte-identical on trunk and the GPU branch. Browser adapter, evaluation and runtime jobs passed in the PR workflow. - Canix’s normal producer refreshed its vendored Fleetix schema and topology; all 50 authored Pkl files validated, and the topology sidecar is current. Fleet-wide routes retain Nomad’s independent media selection and Murph’s oneAPI compute selection. The pinned Toolbelt adapter also passed against modde’s actual local runtime module with host defaults and application overrides.
- The detached-process probe first exposed launcher-exit truncation: bubblewrap exited successfully and killed a still-running descendant. The in-namespace observer now retains that descendant until completion and propagates its nonzero exit. Both exit-0 and exit-7 probes pass.
- A read-only catalogue scan with isolated modde configuration/data found 477 entries and 48 installed entries across Steam, GOG and Epic. This establishes discovery, not live launching. Heroic’s GOG ownership cache was an empty object, so ownership remained unavailable instead of being treated as a complete empty library. Steam account sync and stale Epic ownership remain separate prerequisites.
Outstanding qualification
| Gate | Required evidence or prerequisite |
|---|---|
| Nix regression gate | Approved producer 0d3ce666… passed native package CI, including Nix regressions and packaged lifecycle. A replacement producer revision requires fresh exact-source qualification. |
| GUI-to-Play | Atlas hardware-rendering diagnostics pass outside/inside bubblewrap. Live GUI-to-game rendering, saved configuration and completion evidence remain required. |
| Steam and Heroic | Real exact-install hooks, launcher/cloud-sync setup, native and Wine/Proton/UMU commands, first-run prefix behavior and save continuity. |
| Manager | Live manager forwarding, root lease/marker lifetime and recovery after interruption. |
| Native nested runtimes | Verify supervision, mounts and failure without an unsandboxed fallback for actual Proton/UMU commands. |
| Performance | Release/package runs and retained CSV/configuration/process evidence. Results are report-only. No real-game overhead has been measured. |
| Dependency health | The fixed Cryoglyph graph passed native checks and an audit with zero vulnerabilities. Changed packages and the exact hosted candidate require qualification; four unmaintained and two yanked warnings remain separately tracked. |
| Whole-repository documentation | Badge destinations and Simit policy are repaired. Closeout docs must retain exact-source claims and pass current link/command/mdBook validation. |
| Canix integration rollout | Historical consumer 8240764b… binds the approved producer, actual wrapper, logging/database/manager defaults and packaged lifecycle. Selected successor 3b29bb33… needs coordinator-owned binding and affected checks before final composition, recovery/adoption and activation. |
event-listener 5.4.2 and lru 0.18.2 replace the affected versions on the
dependency edges that accept those patches. Audit warnings are not a clean
security qualification merely because cargo audit returns zero.
For performance, use alternating off/on pairs against one installation/profile, with identical settings and a repeatable scene. The proposed protocol is eight pairs of approximately five-minute runs and 30 seconds of warmup. Logging duration does not terminate the game. Capture/grading requires successful observed exits and at least 100 usable FPS samples and 100 measured frame-time samples after warmup; frame times are never invented from FPS. Report the observed paired deltas and variability without assigning a negligible-overhead verdict.
Flatpak packaging and its host-runtime bridge remain unimplemented follow-up work, outside the required native-Linux launch target.
Repeatable GPU probe and deployment integration
Run the opt-in hardware diagnostic in a desktop session (diagnostic tools are not required by ordinary launches or headless CI):
cargo run --locked -p modde-games --example qualify_gpu -- /absolute/path/to/glxinfo /absolute/path/to/report.json /dev/dri/by-path/pci-0000:03:00.0-render /absolute/path/to/vulkaninfo
The optional final argument also checks that Vulkan enumerates a hardware device first with matching PCI IDs through both boundaries. It does not prove which device a real Vulkan game chooses.
Fleetix’s optional gpu.render.renderNode declares a portable 3D route. Its Pkl
schema owns stable PCI identities and GPU vocabulary; a generated contract and
conformance fixtures bind the Rust parser and Nix projections to that source.
The Pkl evaluator and Tokio are optional for Rust library consumers. modde’s
live device validation, saved installation overrides, sandbox grants and process
evidence remain application-owned.
The shared Rust fleetix::gpu::pci_selector API awaits a crate release; published
Fleetix 0.3.0 does not provide it. modde retains its small runtime parser until
that API is available through a published Cargo dependency.
Canix declares Atlas’s rendering route in host topology and forwards Fleetix’s
purpose-specific gpuRoutes. Toolbelt resolves defaults consistently for
integrated and standalone Home Manager. Its modde-gpu adapter supplies the
application default; browser and mpv media adapters use the resolved media route
and respect Home Manager overrides. Rendering, media and compute remain separate
roles; mpv retains automatic presentation selection and copy-mode decoding.
Scoped GPU changes are published on upstream integration branches before their consuming flake pins and Canix schema/topology sidecars are regenerated. Activation also requires a modde pin containing the runtime module’s GPU option. An older modde module fails the enabled adapter’s compatibility assertion instead of silently ignoring the route. Publishing the shared contract is not live-game qualification.
Performance configuration and durable process evidence now retain GPU routing, PCI IDs, kernel driver, kernel version and immutable Nix graphics-driver roots. Baselines without this provenance require recapture. Paired measurements reject changed provenance and remain report-only.
Publication follow-up, 2026-10-02
Publication of the implementation was authorized on 2026-10-02. The changes are
grouped into host-independent OptiScaler qualification, Rust quality cleanup,
dependency requirements, launch/save/supervision/GPU evidence, Library UI, and
Nix/Home Manager qualification commits. Implementation revision f09ce08
contains the GPU-capable runtime module.
The OptiScaler failure inside Nix came from a test assuming the developer’s
RDNA3 GPU. Configuration now accepts an injected architecture internally, and
the integration test checks RDNA3, RDNA4, and unknown hardware deterministically.
The production entry point still detects live hardware. All-feature workspace
tests, strict all-target/all-feature Clippy, and real containment qualification
passed again. Documentation validation and pinned Simit badge generation pass;
the unsupported ci.check_command was replaced by a required qualification gate.
The committed-source Nix library-regressions gate passed:
- Derivation:
/nix/store/dl9js1argqxia67khy23jj84yb0ysgr5-modde-test-0.7.0.drv. - Output:
/nix/store/gk69xzklyhrl03j1zvkn5r3yxqmx2qxc-modde-test-0.7.0. - Command:
canix cache binary build .#checks.x86_64-linux.library-regressions --include-tests --no-push --max-jobs 1 --cores 2.
Rechecked upstream heads are still open PRs:
- Fleetix
4465108ce8160bdb7b9510211669099ee3a83ac4: PR CI 36934171743 passed. - Toolbelt
9caf679440076addbde1ba9bb0786a29d59cd5f6: PR CI 36935397429 and browser qualification 36935397380 passed, including the real Floorp smoke and exact tab preservation.
Canix commit 56954009d consumes these revisions and producer-regenerated
cloud/publication schema and topology. Topology is current and all 50 Pkl files
validate. The Toolbelt/local-modde module evaluation resolves Atlas’s host
default to /dev/dri/by-path/pci-0000:03:00.0-render, permits an explicit
application override, and retains independent media policy.
These gates qualify implementation contracts. Live games, store/manager integration, real Proton/UMU, GUI-to-game save continuity, packaged paired measurements, and actual game-renderer evidence remain outstanding. Measurements remain report-only.
Published implementation 416b6dd1cd2f93349e0af7c32e12c93c7d2e32f6
is available in Modde PR #3, with
NAR hash sha256-vIn8VOVaaU5uwAElL+8gNPL5XB5uINiUI/p/WVofzbU=.
Canix commit 9d0f1ce46 consumes this revision and its transitive Simit pin.
Full Atlas toplevel evaluation succeeds:
/nix/store/yzsxwr6gpmhlyzmd1hjbmn9cpq05vchf-nixos-system-atlas-26.11.20260823.56c02bc.drv.
The older-module GPU compatibility blocker is resolved; no activation occurred.
The updated native package also builds:
/nix/store/lxb2y07qq4j4b1ykjba696msd4q6lg49-modde-0.7.0.
Synthetic packaged sandbox launches retained detached descendant completion,
aggregate raw exit statuses 0 and 1792 (exit code 7), requested AMD render-node
routing and driver/inventory provenance after private launch-request removal.
Both launches cleared their completion journals. Receipt:
/data/scratch/tmp/opencode/modde-package-gpu-0dvy2lsx/receipt.json.
This is packaged lifecycle evidence, not an actual game-renderer or performance
measurement.
Diagnostics readiness follow-up
The follow-up implementation adds pre-preflight correlated launch records, private supervisor-owned output, GUI tracing, recent-run/log inspection, redacted export, protected retention and separate boundary/inner startup evidence. It restricts automatic raw input-device grants to controllers and requires initialized physical prefixes and a saved Steam Cloud-disabled assertion for managed saves.
The new exact-package fixture can be run against a built native CLI:
cargo xtask library-package-qualify --binary /absolute/package/bin/modde --output /absolute/new-receipt-directory
It uses isolated configuration, SQLite, HOME and data directories and real bubblewrap. It checks successful and detached-failure completion, raw signal status, the private inner-observer channel, retained regular-file game output, secret exclusion from exports, request consumption and cleared journals. It does not establish actual game rendering, store-provider behavior or performance.
Local follow-up checks on 2026-10-02 passed the all-feature workspace suite,
strict all-target/all-feature Clippy, Rust 1.93.0 all-target/all-feature compilation,
minimal and nine-feature builds, containment, diagnostic retention/export and
supervision recovery. The real CLI fixture retained successful status 0,
detached-failure status 1792, signal status 15 and recoverable inner-exec
failure in /data/scratch/tmp/opencode/modde-diagnostics-dev-fixture-20261002-v3/receipt.json.
This is a development-binary fixture, not a qualified Nix package receipt.
Disposable PostgreSQL migration evidence at
/data/scratch/tmp/opencode/modde-postgres-readiness-wdrlNk/receipt.txt reports
modde_fixture|f|t|25: a non-superuser with schema creation rights created all
25 tables. Live Atlas inspection was read-only and reports can|f|t|t|25 for
role, superuser, schema usage/create and table count. No activation occurred.
CI generator and preserved release scope
The latest-CI update selects Simit
1f5a180b7bd8d3cf4f1e8c3a105b33c251fac0e1. Its upstream Pages setup and
concurrency fixes replace Modde’s separate backports. Public flake systems
exclude x86_64-darwin, retired by Nixpkgs 26.11 and already excluded by Modde’s
release policy; supported-system output validation remains required. The retained Modde
compatibility patch preserves release artifact layout, COPR/AUR behavior,
Chocolatey duplicate handling and badge destinations. Windows packaging uses
upstream’s optional Authenticode signing and its four-script ShellCheck coverage.
simit-ci-generator exposes the same patched package used by the development
shell. It is an explicit Nix CI output alongside modde, the library regressions
and Home Manager runtime wrappers. The matrix captures results and uses one
parallel build, one local job and two cores, with 90-minute timeouts and 14-day
artifact retention. Required gates also set Nix limits directly, including the
packaged lifecycle build; matrix limits alone do not govern those jobs.
Aggregate all-feature tests and Clippy are complemented by crate-compatibility:
isolated default-feature tests/Clippy for all nine members, no-default-feature
checks for the six members covered by PR #2, package inventories for its seven
packaged crates, and modde-xtask check. CI and the multi-channel release workflow
are generated separately and both must pass --check --diff. Verification CI
does not publish crates or replace the release workflow’s signing, smoke,
provenance, cache or downstream-publisher contracts.
Feature-specific fixtures declare their prerequisites. The BSA integration suites
require bethesda-archives; the manager subprocess suite requires wow. Additional
minimal-plus-archives, minimal-plus-Linux-integration and minimal-plus-manager
commands retain their execution coverage. A separate minimal CLI command enables
only the game/database prerequisites for export, executable, patcher, scanner and
configuration fixtures. UI game-picker assertions cover both
registered-game normalization and the unsupported-game fallback. These additions
retain the default/minimal member matrix and warnings-denied Clippy.
The actual patched generator passed 107 library tests, 145 focused integration
tests and warnings-denied all-target Clippy locally. ShellCheck was explicitly
present; an earlier apparent pass without the checker is retained as non-covering
evidence. Exact receipts are under
/data/scratch/tmp/opencode/modde-ci-reviewed-20261005-coordinator/.
Lock refresh, generated-file drift, actual Nix generator packaging and
exact-candidate hosted qualification subsequently passed for the merged source
recorded above. These results do not replace the approved 0d3ce666… runtime
producer or qualify later implementation changes.
Hosted containment jobs configure Ubuntu’s unprofiled user-namespace policy
on their disposable runner before testing bubblewrap’s PID/network isolation.
The launch path retains its sandbox preflight and never retries unsandboxed.
Default public flake outputs do not consume private Apple SDK metadata. Authorized
macOS cross-build callers use lib.mkOutputs with explicit macosSdkStorePath,
macosSdkOutputHash and osxSdkVersion; without an SDK, the cross-build package
reports an unavailable prerequisite. Native Linux outputs require no SDK credentials.
Native UI test isolation incident, 2026-10-05
The native UI test command at 19:59:25–20:00:47 UTC inherited the live Modde PostgreSQL configuration. Its temporary data/configuration directories did not override the database backend; fixture constructors honored that configuration, including profile and tool/history cleanup. The successful exit is non-covering for isolated qualification. A read-only inspection at 20:16:18 UTC found one profile and zero rows in the inspected tool, performance and bisect tables. No verified pre-test snapshot establishes the original inventory or quantifies loss.
Tests were stopped and evidence retained in
/data/scratch/tmp/opencode/modde-ci-reviewed-20261005-coordinator/ui-test-isolation-incident.json.
The operator selected continued source repair with production recovery pending.
No live restore or writeback was performed. Backup-service success does not prove
recoverability; privileged backup/WAL inspection and managed recovery remain
separate requirements before production durability acceptance.
The source repair binds UI fixtures directly to SQLite, asserts disposable path ownership and rejects PostgreSQL in the test-cleanup helper before SQL execution. A regression test uses a closed lazy PostgreSQL pool, allowing the refusal contract to be checked without connecting or executing SQL. Native qualification scrubs database/operator environment overrides and uses disposable short HOME, XDG, runtime and temporary paths while retaining the approved Cargo/Rustup caches. Short paths preserve Unix-socket tests’ actual path-length requirements.
The repaired source is committed as
8c02443c047d688448a390ea9109c1be1de41099. All 42 native policy commands
exited zero with unchanged bound Rust sources and all 15 foreign files preserved.
The two no-feature manager commands have no enabled targets; the explicit wow
run passed its 13 subprocess cases. The core cleanup regression exercised SQLite
and the closed-lazy PostgreSQL refusal. A separate UI run passed all 140 library
tests with PostgreSQL environment overrides and that backend feature disabled.
Existing ignored tests remain non-covering. Exact receipts are
native-crate-compatibility-r8/receipt.json, ui-fixture-adversarial/receipt.json
and native-isolation-feature-terminal.json under the retained evidence root.
Earlier managed Simit updates rolled back on admission failures and then exposed Linux-only tools in Apple Silicon’s release shell. Platform-scoped repairs, lock promotion, generated workflow drift and exact hosted package bindings subsequently passed. The isolated source repair and CI closeout do not establish production data recovery.