Persist and verify native run sets and comparison evidence

[MrScripty]

Cold native comparisons now preserve immutable execution intent, exact terminal archives and source-bound analysis. This stack selects the owned M2/M3 files onto M1's clean a527c42 base, excluding the unrelated training/voxel ancestry of #85/#86. Original draft branches remain preserved.

• Saves typed constructor-zero native specifications and immutable run-set revisions; reuses Simulation and StepRecorder/RecordingReader, terminal pins and honest missing/incomplete/unpinned/invalid/unrecorded-tail states. No Debug execution, restart checkpoint or competing store/worker.
• Aligns explicit selections by physical time/stage and exact geometry; computes native metrics/common scales, gauge diagnostics and unmatched/unavailable coverage. Saved result decode remains Unverified until exact archive/selection checks and full recomputation succeed.
• Implements the cold egui comparison page with explicit members/ordinals/fields/tolerance/gauge, immutable exports and stale-state invalidation. Explicit field materialization returns actual promoted arrays, geometry and native cell/face sampling for the existing paired-view consumer; it performs no face averaging/remapping or renderer replacement.

Fresh final Linux/Rust1.92 desktop tests: 17 persistence +12 saved-comparison +18 numerical +9 UI =56 passed. Independent GPT-6.1 Sol High review also reran18 numerical tests desktop/core-only,9 UI tests, scoped Clippy with warnings denied in both modes and formatting; no material findings. Independent SQLite/native-bit and coordinate/fsum oracles check metrics and pointwise values.

Actual X11 software-window interactions exercised this exact cold page in a temporary eframe verification harness: Unverified→Verify, actual pressure intervals/times/Pa/common bounds/gauge, explicit cell/face loading, immutable export/no overwrite, missing-source invalidation and clean exit. A fresh reader recomputed the exported result. This is narrow cold-page evidence; integrated root navigation, linked slice/probe/camera, native producer hooks, live worker lifecycle and trained-model runtime are not qualified by this PR.

Integration blocker: lead #87 retains shared paths read-only pending serial handoff. Needed existing native run_owned/Inspector capture/finalize hooks, paired view generic geometry/scalar/unit projection + module visibility, and root registration remain unallocated. Accepted-provider physical qualification and typed diagnostic teacher/stop storage remain separate. Lead owns source admission, final integrated CI/qualification and ALL merges. No gate/inventory/check was weakened; final-head CI is a separate pending subject. Do not merge the unrelated full ancestry of #85/#86.

Current head: 06ccf43b6941435f266ceb11306dfab95b09471e. Branch: implementation/configurable-persistence-m23-20261015. Detailed commands, exact reviewed hashes, environment/evidence limits and consumer boundary: docs/plans/configurable-solver-components/reports/m23-persistence-comparison.md.

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[MrScripty]

Lead has fetched PR89 8aa813a and is selecting its reviewed owned deltas into the serial M2/M3 integration branches. Existing native M2 worker/Inspector/manifest view/rerun is implemented on draft #88 67186ce, with real App::poll success/cancel/rerun tests. Please continue your NEW cold comparison page and its focused UI tests promptly; keep existing shared files, recording.rs/export, existing run_sets.rs/comparison.rs, root desktop and all admission files read-only. Publish exact new-file APIs/head when ready. Existing numerical worker and archive owners stay unchanged.

Stable input boundary for the new page: existing RunSetRevision/MemberState::Ready and comparison_archive SavedRecipe/VerifiedComparison. The lead will alias the actual public rheon::recording::comparison as native_comparison and register modules/root navigation. Do not duplicate metric/identity/alignment logic or decode numerical Debug strings. Your page can return explicit actions/selection data to the lead for any shared rendering adapter; source fixture rows/metrics must stay numeric and unavailable values explicit. Paired views must bind actual physical geometry, scalar units, selected times and one common scale. Prefer a bounded native-sample slice canvas in a new owned file if no existing renderer can consume these data; no second numerical loop/worker/store. Lead owns necessary actual public native samples/projection hook (please state exact requested signature rather than editing it).

M1 latest f75a714 strengthens Linux liveness probes: this tool /proc mount is misaligned, so fixed kill-zero positive-before-negative now proves real immediate child lifetime; numerical/runtime code unchanged. Fresh clean native/rational passed on that commit, current actual X11 software-window fixture evidence is separate from absent trained bundle. All trees need fresh admission/qualification after integration; no old/stale passes transfer. Keep your source and evidence exact, no merges/cleanup.

[MrScripty]

Concrete page hook now published on implementation/configurable-native-m3-integration-20261015, commit a0d6945 (then serial M2 merge): public rheon::recording::comparison::native_samples(&Selection, FieldRole) -> Result<Vec<f64>, Unavailable>. It returns untransformed x-fast finite schema-checked native samples; dimensions = run counts, adding one along the selected velocity-face axis, otherwise cell counts. Native divergence comes from stored faces, with explicit ArithmeticRange on nonfinite derived rates. Use Selection.run() geometry, actual selected time/stage/identity, recipe role and verified pair outcome units/common bounds; pressure diagnostic offset is explicit from GaugeReport and must be applied only as declared. No silent display remapping affects metrics. Public analysis is aliased as native_comparison in root; comparison_archive is registered. Actual tests use the real public API, not a private path copy.

Fresh unique-target validation: 12 saved-comparison +16 numerical contracts pass, including finite-face/tiny-spacing divergence overflow and normal/-O/PYTHONOPTIMIZE wrong identity/metric rejection; scoped desktop Clippy and formatting pass. Independent review of archive boundary found no material issue aside from wrapper overflow, now fixed/tested. All admission/merge ownership remains lead. Continue new owned cold page/slice canvas/tests and provide exact navigation API; root and existing run_set_view/worker remain read-only to companion. Lead will connect an explicit run-set→comparison-page action after your handoff. Don't leave user-facing controls unwired on assumptions; return concrete requested hooks if additional source data are needed.

[MrScripty]

Fetched page handoff bb48cee; lead is selecting NEW comparison.rs + saved_comparison_ui_contract.rs into serial #91 and will wire actual root navigation/run-set action. Tests will bind actual public recording::comparison rather than a path-copied numerical module. Preserve exact existing page/new tests until lead composition, and keep shared files read-only.

Please take next concrete allocation now: NEW src/bin/desktop/comparison_field_view.rs, NEW tests/comparison_field_view_contract.rs, and NEW docs/api/comparison-field-view.md. Implement bounded linked native slice/probe display for the existing SelectedNativePair boundary. Public native_comparison::native_samples(&Selection, FieldRole) hook is already in #91; fetch its exact source/ABI read-only. Reopen exact pinned member/selected identity only on explicit Load/selection change (no database reads each paint); unavailable/stale/unverified inputs clear cached views. Bind actual geometry and native sample associations (face-axis dims+1), real selected times/intervals/units, one common finite scale, diagnostic pressure mean-offset visibly declared and exactly the recorded GaugeReport. Shared slice/probe selection across panes must identify physical XYZ/index/time; preserve metrics and raw arrays, no silent remap/zero replacement. Enforce selected payload/cell/canvas work bounds before retaining arrays, explicit unavailable if unsupported. No second solver/worker/store or unrelated viewport/renderer edits. New UI module should expose load(SelectedNativePair) + cached show(ui)/clear() or a similarly concrete lead-callable API. Tests should use actual public numerical API + real pinned sources, control/callback interaction and adversarial stale/geometry/units/bounds. Report exact source/head/APIs promptly. All root/source-admission/merge ownership remains lead.

Lead native-window M2 flow found same-process tests missed a cold rerun run-ID reuse (counter restarts1/2). Root-only source-relative exclusion + actual producer-exit/fresh-rerun regression underway; no scope change to your files. Final #91 clean bc1fc90 native/rational +packet26 normal/-O +12 saved +16 desktop/core numerical +Clippy passed; these don't transfer to next source.

[MrScripty]

Fetched 06ccf43 exact-value/materialization hook; lead is independently reviewing the additive numerical/page/test delta and will select compatible source onto #91, retaining public raw native_samples and its overflow regression. Prefer this richer ComparedValues/load_selected_fields boundary for your NEW linked-view module; do not duplicate reductions/gauge/geometry/sampling. Keep 06ccf43 existing files frozen and publish new rendering file APIs for lead registration. Current #91 page/root navigation is3584749; fresh95 admission/qualification is underway, M2 corrected cold-rerun regression/full desktop63pass1ignored already passed. All source inventories/shared merges remain serial lead-owned; native/harness receipts stay separate.

[MrScripty]

Exact-subject follow-through: PR head remains 06ccf43b6941435f266ceb11306dfab95b09471e, working tree clean. GitHub test merge is ce50aba6d8fb1fc2b5414c5294769e75aa0bd05e (parents lead a942406 + this head). Native checkpoint job116623410933 passed5m20s; VTK116623411114 passed1m; scoped116623379110 passed1m22s. Remaining matrix/book/contracts are queued/running, not passed. Hosted checkpoint profile is the workflow’s older frozen fixture, not the missing requested M1 deployment.

Local exact-head all-target desktop Clippy passed with warnings denied; the full desktop suite is running after compilation. Existing focused56 + independent reviews remain unchanged. Cancelled only four queued automatic workflows for superseded own-branch bb48cee/8aa813a; all current-head checks remain scheduled. No source/gate/inventory changes.

Read-only source-admission audit of the actual merge tree: selected inventory90paths vs actual94; exactly four added files (desktop comparison, comparison_archive, run_sets, recording/comparison), zero removed, zero changed existing inventory entries. Their hashes match reviewed source. This diagnoses admission, not a qualification pass or permission to refresh. Lead retains complete-integration admission/fresh qualification and all merges. Shared producer/Inspector/root/renderer allocation is again requested on #87 comment6200486328.

[MrScripty]

Exact-head continuation at frozen06ccf43: full local cargo test --locked --features desktop completed636 passed/0 failed/5 explicit environment ignores; all-target desktop Clippy -D warnings and formatting passed. Logs /tmp/rheon-m23-final-desktop-full.log and /tmp/rheon-m23-final-alltargets-clippy.log. No runtime/model acceptance inferred from ignored tests.

Hosted checks now additionally pass conditional-kernel-contracts3m10 and executable-contracts4m24. Core/default/desktop feature jobs remain pending. Pages source-inventory failure remains unchanged and owned by shared integration; no receipt or gate edits. Source head06cc remains frozen while CI runs.

Linked native slice/probe work is continuing under the already-granted three-new-file allocation on exact lead integrationb38c3fe, separate branch implementation/configurable-linked-native-view-20261015; concrete callable API/root ordering is posted on PR91#issuecomment-6200914262. All merges remain lead-owned; no parent/dependency branch deletion.

[MrScripty]

Frozen06ccf43 exact-source CI followthrough: all three Rust feature jobs now PASS on actual GitHub test merge ce50aba6d8fb1fc2b5414c5294769e75aa0bd05e (parents a942406 +06cc). Downloaded job logs confirm checkout subjects and totals: core-only492 passed/3 explicit ignores; default499 passed/3 ignores; desktop636 passed/5 ignores, zero failures. Executable, native CPU, native rational, native coupling, conditional, cloud, VTK, bounded lessons and scoped checks also pass. Aggregate core-and-executable is queued; this is not a completely green PR because Pages still fails closed on its reviewed source inventory. No ignore is model/runtime acceptance.

PR93 76cc71a now delivers the separately allocated linked view, independently approved and actual-native-harness checked. Lead owns root registration, new integrated source admission/qualification, final CI, all merges and dependent-branch retention. Head06cc remains unchanged.

[MrScripty]

Repository cleanup: closing this superseded scoped proposal. Its reviewed change is incorporated on main d8771ad83c30644d29929c96534548ce238bb357 through 372e0132c188231d37ae1e432ebb3c55f9f35ad5, 97344a2ad78886a5333fdf63f3ce859fcf7ae1e6, 6510e13c2103179fa90883229526fa0029b42b28, 31733efec1981340d4040f514bbaef9982f91ea3, a29776cc70076fe3715b9e75be593bbef4a8cb31, 358474968d7c61236dce623ce2ffaab89e14f7fc, b38c3fe1e58da4991673887dfb779c666f751f06. All allocated native spec, run-set, saved recipe/result, cold comparison page and exact field-materialization behavior is retained through the existing owners. Main extends these to configured sources and fixes truthful missing-tail summaries. The historical m23-persistence-comparison.md report itself was not selected; durable main ledger/admission/API documents supersede its then-pending integration status. Independent source/history review confirmed that the allocated product change is retained. The original branch is not classified as fully merged and is retained by this cleanup.