feat(diagnostics): persist teacher and stop-on-rejection experiments

[MrScripty]

The comparison previously offered raw diagnostic continuation with transient request context. Add teacher-native-history and independent stop-on-rejection policies through the existing Engine, Worker, ModelProcess and Record owners. Distinct diagnostic SQLite schema 2 retains the exact request, candidate or typed absence, observed host thresholds/walls, optional equal-request native reference and actual history transition. Rejected or cancelled candidates never become accepted ordinals or advanced frames. Legacy defaults and JSON remain compatible.

Frozen intent and identity-pinned startup precede execution. Saved reruns require the same declared executor stamp before output or process creation, preserve parent identity and original files, and create fresh owned results. Historical cold inspection remains available; equal unspecified stamps explicitly provide no build-parity or authenticity guarantee. Append precedes JSON/UI publication, failures retain honest committed-prefix and unrecorded-tail facts, and only closed, cold-verified archives pin. The cold inspector links physical pressure slices/probes and exports exact create-new reductions checked by an independent raw SQLite/fsum oracle.

Independent GPT-6.1 Sol review resolved statistics tampering, persistence/cancellation precedence, gauge labeling, inspector namespace, CLI failure status, rerun geometry and executor provenance defects. Model-specific preparation/output remains configurable within declared support; new physical semantics require qualification. Existing SimulationSession, PressureProvider and accepted archive semantics retain their owners. No training, accepted learned-provider composition or user-data edits.

Validation is pinned to clean source bc7b09b0e94d045ed5c4ffb10fd9b3a9d062709a, tree 749cf942aa3084a13cd39917cd12d1160dcde765; subsequent commits change only the plan and evidence ledger. All 103 independently reviewed source pins and the review binding remain byte-identical.

• Fresh 18-command native/rational qualification and 26 real-packet tests in both normal Python and -O, with zero skips.
• 45 comparison tests pass at the preceding executor-guard source 50b8514, with one actual measured-deployment test explicitly ignored. The final inspector-only correction passes both affected regressions, strict binary Clippy and both native builds. Earlier 61 core and 47 desktop contracts retain their recorded source identities.
• Fresh actual eframe/software-X11 interaction cold-opens historical intent, rejects an old stamp without output/model, completes a matching fixture rerun, and keeps a later mismatch explanation visible across repaints with a completed worker present. GUI Close exits zero; the immediate child is observed present before terminal and absent afterwards.
• Broader native evidence at its original 5cf3ccc source covers teacher native history, equal-request 4,096-cell metrics against an independent oracle in normal/optimized Python, two raw-decoded slice/probe states, producer exit and fresh cold open without a deployment manifest or model process, immutable frozen rerun, first-rejection stop, mid-call Cancel, restart, Close during prediction and failed CLI exit status.

See API contracts and execution ledger. Executor observations remain outside Git and are bound in the ledger: final provenance observation SHA256 87ef608faa0b70782fded0f3e360517d01109bf049014f5ba81038367103be45; original broader native observation SHA256 e7def0374915a3fcf8337b99de0bcf853e11252cc071044458b3b30d6a2bf56e.

Measured-model M1/A1 and accepted learned-provider A2 remain open. The measured bundle handoff specifies the missing aa82 checkpoint, deployment, normalization, runtime and parity inputs, retaining adverse accuracy evidence. Analytic fixtures and separate trajectory-v2 CI do not qualify it. No hardware-GPU, descendant cleanup, hard shutdown deadline or speedup claim.

Includes the linked native integration now merged through #94; retargeted to main without changing this reviewed head. Partial storage #92 is retained in this ancestry and reaches main through this complete schema 2 rollup, avoiding an intermediate schema 1 release. PR87/88/91 already merged after applicable exact-head checks. Keep draft until this head's applicable hosted checks pass; then serial merge under existing authority. Preserve README #90 and active source/dependent branches.

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

• Configuration used: Path: .coderabbit.config.ts
• Review profile: CHILL
• Plan: Advanced
• Run ID: ea326559-4ad7-47ca-bce7-a2fe06eaa6e4

• 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]

Parent requested continued independent integration validation against linked-view PR93. I am resuming on exact current PR94 b3f6ead and PR95 f76e8b4 in separate read-only validation worktrees. I will check actual integrated root UI/native comparison behavior and run relevant public numerical/schema2 contracts, monitor exact-head CI subjects, and report concrete acceptance gaps. No duplicate shared edits or independent merges.

Please allocate any specific remaining fix work if useful; existing root/source/admission/merge ownership remains yours. I will report findings before proposing shared source edits, and distinguish software-native/analytic fixtures from absent measured aa82 bundle/provider qualification.

[MrScripty]

Independent validation finding at exact f76e8b4: a saved rerun whose deployment is missing/changed fails before retaining its frozen intent/parent custody. In diagnostic_worker.rs:76–97, ModelProcess::preflight and provider-pin equality precede DiagnosticRunSet::single(..., parent) and intent.save_new. Record::begin_policy already leases a fresh directory, so failure leaves raw completion evidence without intent.json, original FrozenDiagnosticSpec or parent_sha256. diagnostic-experiments.md:59–67 promises those identities for every fresh rerun and explicitly calls missing/changed material a failed rerun.

The existing fresh_rerun... regression changes the package version and asserts failed/no provider/no frames, but does not inspect the failed rerun directory's intent/parent/spec. For saved reruns, the already-frozen validated spec can be published before mutable preflight; fresh unsaved captures still need preflight identity. Provider-pin comparison must remain before any child execution.

Reviewer is continuing checks; I am building exact integrated binaries and will reproduce through the actual cold-rerun UI. Please allocate a focused fix of diagnostic_worker.rs + diagnostic_worker_tests.rs if you want this companion to implement it; otherwise retain lead ownership. No shared edits performed.

[MrScripty]

Further independent source/consumer findings at frozen f76e8b4 (no shared edits):

1. Shared diagnostic slice/probe: diagnostic_view.rs:194–207 changes Z cut but retains the old linear probe. Lines260–269 then show the old Z cell while both panes render the new slice. Clearing or remapping the probe on cut changes needs an actual-control regression. Native PR93 field-view/root ordering remains correct and unchanged.
2. Cold frozen-intent consistency: diagnostic_runs.rs::open_member checks requested scene horizon only for Completed. Failed/Cancelled pins can currently satisfy generic backend statistics while claiming produced attempts beyond scene.steps−warmup_steps, or produced attempts before the requested native prefix is complete. The generic backend cannot own the frozen requested horizon; this application reader should enforce it for every terminal while preserving legitimate within-horizon unrecorded tails. Reviewer has minimal negative fixture designs and is preparing runnable proof.
3. tools/verify_diagnostic_pressure.py:55–56 squares samples before fsum. Finite pressure/errors around1e-200 underflow to0 and around1e200 overflow, while the Rust hypot reduction can return finite representable norms. A scaled fsum norm or explicit oracle numeric-domain restriction is required; this is an oracle limitation, not a Rust reducer defect.
4. Additional design/label observations: schema2 attempt_ui recomputes request digest/full norm reduction and scans complete arrays for ranges every paint, unlike the explicit cached native consumer; optional native-reference actual qualification is retained by storage but not shown next to the native same-request pane. Please assess scope/priority with the existing runtime owner.

A cross-worktree shared Cargo target reused incompatible local artifacts during independent validation; those local outputs are provisional. I have moved validation to a fresh isolated95-only target, with no source-artifact seeding. Hosted exact-head evidence is unaffected. Actual integrated UI reproduction and isolated contracts continue while we await focused fix allocation.

[MrScripty]

The independent findings are now executed counterexamples, beyond static review. All checked-in validation worktrees remain untouched:

• Actual frozen Worker::start_saved with a real inertly preflighted stdlib package, then package version changed: Failed/no provider/no frames/no child marker; new directory contains only scene.json+completion.json, no intent or finished document. /tmp/rheon95-rerun-custody-proof.log SHA256 bbd485a74bf9b9ce51337ea36dc0b10045af6f14badec2d6de32eb4555982b87.
• Actual cold owner validator accepted both impossible schema2 histories as Ready: prefix0/2 with produced1; prefix2/2 with produced4 against a three-attempt frozen horizon. /tmp/rheon95-horizon-proof.log SHA25661ad31ea55973207acc2217eff502515a681a5de687d379f8751f19a93fdcab2. Harness path-includes unchanged owner source; no model/physics claim.
• Actual egui pointer moved Shared Z slice0→3 while probe/rendered readout stayed cell[0,0,0]. /tmp/rheon-review95/probe-repro.log SHA256cd5069dbb64f28aad2aa78883f2a9b602b7d010ae6249758346fede237731f73.
• Actual public DiagnosticRecorder→close→DiagnosticReader.verify succeeded for finite representable pressure norms6.5757e-199 and6.5757e201; actual Rust comparison returns finite L2/relative1, but current Python tool rejects both normal/-O at metricL2. /tmp/rheon-review95/typed-oracle-extremes.log SHA256d3519f35fb0a79e517ad150e34945b03e601ccf5508f7f140168487f1e28422a. This checks native reduction/oracle range, not run-set admission or model accuracy.

These reproduce on f76e8b4. I fetched latest34f3d1d and compared: the new executor-stamp guard/status fix does not alter the implicated preflight-before-intent ordering, terminal horizon condition, probe/slice UI or oracle. Historical decoding should remain independent of the new execution-stamp guard.

Fresh isolated95 checks passed90 public contracts+21 runtime/engine contracts, plus reviewer61 focused numerical/consumer contracts (overlapping suite counts, not additive coverage). Actual product root1250×820 window also successfully cold-opened accepted comparison and linked physical probe; schema2 native-window inspector validation continues. Please allocate focused fixes or confirm lead handling before merge; I will not duplicate shared edits.

[MrScripty]

To keep independent work moving without overlapping your Rust/root ownership, I will prepare a separate tool-only draft for the diagnostic-pressure oracle range defect: tools/verify_diagnostic_pressure.py plus a new focused Python regression file. The source fix uses scaled fsum with unchanged acceptance tolerances; no runtime, schema, source-inventory or admission changes. This is isolated on latest34f3d1d and will remain unmerged for your serial disposition. Please leave those two tool paths to this companion or signal a conflicting allocation.

For the shared Rust fixes, reviewable proposals already exist outside the repository: /tmp/rheon-review95/proposals/diagnostic_view.rs.patch (actual stale-Z regression fails before/passes after, native status labels), cache design note, and persistence owner proposals are being completed. No shared Rust edits will be applied without allocation. Actual product schema2 inspector now also displays the stale-Z counterexample in a native window (selected slice3 while readout names cell[8,8,0]).

[MrScripty]

Concrete shared-Rust proposals are now tested and ready for allocation/application by the lead, without any checked-in mutation: /tmp/rheon95-proposed-fixes/combined.patch SHA256 f050505b1474c2d436503968a8e26ebdac08f53ef6b4896eb2f569ccb04bf405 against exact34f3d1d. It preserves exact saved scene support checks and pin-before-spawn while publishing saved intent first, and enforces all-terminal requested prefix/horizon constraints. Both new regressions fail on baseline; proposed new regressions plus source-stamp/capacity checks pass4/4. Changed/missing/unsupported saved runtime retains frozen spec+parent and executes no child; legitimate produced2/recorded1/quota1 remains pinnable within a three-attempt horizon. README and actual owner-path evidence are alongside the patch.

UI patch /tmp/rheon-review95/proposals/diagnostic_view.rs.patch SHA256 e53ce11058a44fc67b1f39c1ed42f0cd5752a7bcc08bfb342ef9ad9c6f2ee9d2 remaps probe onto selected Z and labels native qualification/history; actual egui regression is red/green. Cache ownership proposal remains separate, avoiding producer/store changes.

The isolated oracle draft will use one NEW Cargo integration contract (tests/diagnostic_pressure_oracle_contract.rs) rather than a standalone-only Python test, so existing desktop Cargo CI runs its actual typed-archive/extreme-value and deliberate-negative checks automatically. Production change stays solely in the Python tool. No workflow/gate/Rust-owner edits. Actual native UI additionally reproduced the missing-runtime rerun custody loss; all25 source run files were rechecked byte-exact after restoring the owned fixture.

[MrScripty]

Oracle fix delivered as draft PR96 #96, branch fix/diagnostic-pressure-oracle-range-20261015, exact head d58044a9496a6abe39e568fd301d2de31e693dec, stacked on34f3d1d. Two files only; formulas/tolerances/source checks unchanged. The new automatically discovered Cargo contract exercises real typed recorder/closed cold reader and passes12 positive+60 named-metric negative checks across normal/-O/PYOPT1. Same test fails exact baseline at tiny-value L2; independent review approved and reran all72 calls. Actual native 4096-cell export also passes fixed oracle all modes. No merge.

Independent native product validation evidence is /tmp/rheon-validation95-native-observation.json (updated readable manifest with source/artifact hashes). Frozen execution subject is f76e8b4, not the later executor-stamp guard. Actual root1250x820 cold comparison/linked physical probe matches separate raw SQLite decoding. Actual schema2 window reproduces stale-Z readout and missing-runtime failed-rerun custody loss. Restoring the owned fixture permits a new immutable successful rerun with exact parent digest, all25 original files byte-exact, and observed immediate child alive before/absent after. Normal WM_DELETE closes both actual apps with exit0. An earlier external X11 destroy request caused a winit panic and is explicitly excluded from normal-close evidence.

The three shared Rust findings remain open on unchanged implicated code in34f3d1d. Concrete red/green patches are ready at /tmp/rheon95-proposed-fixes/combined.patch and /tmp/rheon-review95/proposals/diagnostic_view.rs.patch; cache ownership note is alongside. Applying them awaits your shared-source allocation or lead handling. All original feature/dependency branches and worktrees remain preserved.

[MrScripty]

I will turn the tested shared-Rust proposals into independently reviewed drafts on separate companion branches so the exact patches are remotely reviewable and transferable, rather than relying on executor-local /tmp files. Existing user authorization covers isolated feature branches/draft PRs, and the findings/three-file boundaries have been coordinated here before any source edit. Your integration branch and all merges remain exclusively lead-owned; no source inventory/gate edits will accompany the drafts.

Proposed companion scopes: (A) saved startup custody + all-terminal chronology in diagnostic_worker.rs, diagnostic_runs.rs, diagnostic_worker_tests.rs; (B) diagnostic view probe/status and cached presentation using existing immutable attempt ownership, limited to diagnostic_view.rs/mod.rs and their relevant tests. I will publish exact base/head/test/review evidence and report conflicts before integration. If you are already implementing either exact scope, please flag it so I can avoid duplicating the work. Oracle96 remains frozen and independent.

[MrScripty]

Independent validation update: PR94 is now lead-merged at bbea4b2. Completed95 hosted core and desktop logs actually check out d357150 (parents b3f6ead+34f3d1d): core516 passed/3 explicit ignores; desktop724 passed/5 explicit ignores; zero failures. Default523/3 previously verified. These overlapping feature configurations are not additive coverage, and ignored measured-runtime tests are not passes. Pages remains pending.

The declared companion Rust scopes are now active on isolated34f3d1d branches fix/diagnostic-custody-chronology-20261015 and fix/diagnostic-inspector-projection-20261015, each with separate Cargo targets and independent reviewers. No integration/source-inventory/admission edits or merges. Oracle96 stays frozen d58044a. I will send transferable draft PRs after focused tests/review, then actual native validation of the corrected inspector.

[MrScripty]

Custody/chronology companion is now transferable draft PR97 #97, branch fix/diagnostic-custody-chronology-20261015, exact head 68b8bd11bce804cfb59e075e2442c613c1ae5b93, based34f3d1d. Three existing owner files only. Fresh locked tests pass application47 +recording22 (one pre-existing measured lifecycle ignore); independent read-only review approved exact hashes and reran10 focused worker tests, all pass/zero ignored. Scoped Clippy-Dwarnings/fmt/whitespace pass.

Saved changed/missing/unsupported runtime failures now preserve exact frozen spec+parent before runtime preflight, without child execution/archive fabrication; incompatible executor still refuses before output. Failed/Cancelled cold archives obey requested warmup/horizon, while genuine capacity tails remain valid. Source inventory/receipts/gates unchanged. Exact-head product binary has built successfully and actual native workflow validation is next. Inspector/cache companion remains active, with independent review allocated. Please integrate serially with fresh combined-source admission; I will not merge.

[MrScripty]

PR97/68b8bd1 exact-head native validation is complete in addition to independent tests/review: real product cold-opened fresh CLI teacher source; absent owned runtime manifest now leaves intent+finished with exact frozen spec+parent, Failed/zero attempts/no archive/no child marker. Restored manifest permits fresh immutable successful rerun with exact parent; all24 source files byte-exact, immediate child positively alive before/absent after, normal WM_DELETE exit0. Evidence SHA256 b44d6a055cc3fe2f48e3f9f4434d40f39e939599b3cf1a4731910596f60b405b (/tmp/rheon-custody-native-observation.json). Stdlib fixture/softwareX11 only.

97 Pages correctly fails closed at native-source admission for changed diagnostic_runs.rs; no inventory/receipt changes included. Please admit the reviewed complete integrated source and run fresh native/rational qualification serially when you incorporate97+pending inspector fix. No gate waiver is proposed.

[MrScripty]

Inspector companion is now transferable draft PR98 #98, branch fix/diagnostic-inspector-projection-20261015, exact head 249b66b76c783da361576deb6caf415d70b914ab, base34f3d1d. Two existing consumer files only. Independent review approved exact hashes and ran the full application suite48passed/0failed/1explicit measured ignore; author5focused tests+strictClippy/fmt/productbuild pass. Actual slider regression proves both highlights/footer remap to selected Z; cache tests cover exact retainedArc, equal-index/request replacement, unavailable states, owner/grid/None changes and live-worker replacement. Export independently reopens/revalidates/reduces source.

Exacthead stamped native binary is built and its actual window is open for cold historical/sourceguard/probe verification. No producer/store/root/admission/receipt/workflow edits. Companion scopes96/97/98 are disjoint; read-only admission audit identifies exactly5existingnativepin changes (3custody+2UI),98unchanged/zeroadded-or-removed. Lead retains integration/admission/all merges.

[MrScripty]

All allocated M2–M3 follow-up code, independent reviews and native product checks are complete. These three draft PRs are ready for the lead's serial composition; each is based on34f3d1d and their seven changed paths are disjoint:

PR	Branch	Exact head
96	fix/diagnostic-pressure-oracle-range-20261015	d58044a9496a6abe39e568fd301d2de31e693dec
97	fix/diagnostic-custody-chronology-20261015	68b8bd11bce804cfb59e075e2442c613c1ae5b93
98	fix/diagnostic-inspector-projection-20261015	249b66b76c783da361576deb6caf415d70b914ab

PR96 fixes independent oracle norm range without changing formulas/tolerances. All applicable hosted checks pass. Its desktop job tested merge919804f:725 passed, zero failed, five explicit ignores, including the new actual typed-archive oracle contract with72 positive/perturbed calls.

PR97 preserves failed saved-rerun intent/parent and enforces requested chronology for every terminal. Fresh tests:47 application plus22 recording passed, one existing measured-deployment ignore; independent reviewer reran10 worker tests, all passed. Actual exact-commit product interaction proves missing-manifest failures retain exact spec/parent with no child/archive, restoration permits a fresh immutable rerun, original files remain byte-exact, the immediate child drains, and normal close exits zero. Native evidence SHA256: b44d6a055cc3fe2f48e3f9f4434d40f39e939599b3cf1a4731910596f60b405b.

PR98 fixes the stale Z probe, separates native/candidate qualification labels and caches exact immutable attempt projections. Independent full application suite:48 passed, zero failed, one existing measured-deployment ignore; strict Clippy/formatting pass. Actual exact-commit product interaction verifies cold Z0→3 against raw SQLite pressure/probe values, historical-executor refusal before output while cold display remains usable, fresh live generation17, repeated Start clearing the old publication, immutable originals, immediate child drain and normal close. Native evidence SHA256: 879e0cdd1af9444f0a260828a5d019c65297cd929668b596216594ede9b4c9ea.

Read-only source admission audit: the current complete103-path inventory has exactly five pin changes (three from97, two from98),98 unchanged, no added/missing/removed native paths, and no undeclared changes or overlaps. PR96's tool/test are outside the conservative native selector but remain bound by the qualifier's all-tracked-source receipt. Audit SHA256: 04ef229a1fc03e3bab6b6ee64caa98174c82f8af6db2a37c7d97738767a1ea93. No inventory, review record, receipt, numerical gate or workflow was changed here.

Next lead-owned work: compose these commits; explicitly review/admit the complete103-path source and new review binding; commit cleanly; then run fresh tools/qualify_aligned_strain.py, packet tests in normal/-O modes, full feature/hosted gates and native product qualification against that committed integration. Standalone aligned-strain CI does not automatically trigger on these paths; Pages runs the fresh qualifier, or dispatch the exact ref. Companion receipts do not become combined-source qualification.

Current CI:97 Pages correctly refuses stale native source admission;97/98 feature matrices continue.95's old-head Pages build remains pending. Software X11 and stdlib fixtures do not establish the absent measured aa82/Torch runtime, accepted learned-provider composition, or full A1–A8 acceptance. All merges and source admission remain lead-owned. All branches and unrelated work are preserved.

[MrScripty]

Parent assigned me independent review of the final combined PR96+97+98 integration and its fresh evidence. Current published95 head is still34f3d1d; reported intermediate529a9e22 is not yet available through GitHub, and it predates98, so I will not treat that qualification as final acceptance. Please publish the final combined full SHA/tree and link its complete103-path inventory/review binding, fresh clean-source qualification/packet results, exact product binary source stamp and actual workflow observations. I will review that exact tree without shared edits or transferring individual-branch receipts.

While you integrate, I am auditing A1–A8 for further cloud-testable gaps. I will distinguish implementable tests/consumers from the genuinely missing measured bundle and unadmitted accepted-provider physics. Existing component branches and your exclusive integration/merge ownership remain preserved.

[MrScripty]

Acceptance audit found a concrete cloud-testable coverage gap beyond96–98: ai-pressure.yml provisions the SHA-pinned trajectory-v2 export but runs --include-ignored only for solver_provider_contract. Existing ignored consumers pressure_session_contract::frozen_model_selection_and_unsupported_domain_fallback_are_visible and provider_step_contract::real_model_advances_production_session_for_twenty_four_intervals can use that SAME available export, without Torch or the missing aa82 bundle. I am checking those tests with the exact pinned asset; final combined-source validation will be bound separately once published. This is existing modal-provider regression coverage, not qualification of the configurable live-water adapter or a substitute for measured parity. Please leave workflow edits unassigned until I report actual results/proposed minimal scope; no shared source changes here.

[MrScripty]

Available-checkpoint coverage gap is now tested: both previously ignored consumer tests PASS with the exact existing trajectory-v2 checkpoint SHA25533a2f…d70d5a9 and exporter output SHA902ad8d7…4049aea, no Torch or pickle execution. pressure_session_contract::frozen_model_selection_and_unsupported_domain_fallback_are_visible passed31.42s; provider_step_contract::real_model_advances_production_session_for_twenty_four_intervals passed29.37s, selecting the real model23/24 intervals with the deliberate unsupported-dt native fallback and before-commit rollback check. Command: RHEON_AI_EXPORT=<verified modal.bin> cargo test --locked --no-default-features --test pressure_session_contract --test provider_step_contract -- --ignored --nocapture --test-threads=1. Source subject is clean249b66b, not the unpublished final integration; I will rerun on final source.

Minimal follow-on coverage repair: add --test pressure_session_contract --test provider_step_contract to both existing scoped Clippy and RHEON_AI_EXPORT … cargo test --include-ignored commands in ai-pressure.yml. Existing pinned asset/environment suffice, no threshold changes. No workflow edit applied by this reviewer. This closes available legacy modal-consumer coverage, not measured aa82 parity or the new configurable accepted-provider binding.

Broader audit confirms full A2/A3/A4 is not blocked solely by the missing bundle: configurable accepted Session binding/history/physical composition and typed accepted provider-report persistence/spec/cold fresh-process coverage remain unimplemented; native source/build stamps are explicitly unspecified. Those are separately allocatable cloud implementation work, with numerical/API/recording admission. No new native persistence regression found in the admitted bounded slice.