simjecture 0.5.2rc2__tar.gz → 0.5.3__tar.gz

{simjecture-0.5.2rc2 → simjecture-0.5.3}/.gitignore

@@ -20,6 +20,8 @@ artifacts/
 !demos/agent_roles_validation/record/workspace/artifacts/
 !demos/agent_roles_validation/record/workspace/artifacts/*.json
 data/
+!src/conjecture_solver/llm_bench/data/
+!src/conjecture_solver/llm_bench/data/*
 .private/
 
 integrations/dsh/node_modules/

simjecture-0.5.3/AUDIT_LOG.md

@@ -0,0 +1,164 @@
+# rc3 audit and polish log
+
+## Scope and safety
+- Requested 2026-10-01: audit code/performance, GUI usability and design, wording, documentation, public API, contributions, and long-running harness behavior.
+- Branch: `audit/rc3-polish`; baseline: `2f66b8d06c20b18cfea20188acad60b83fa25ec0` (`0.5.3rc3`). No main merge or history rewriting.
+- Changes are committed in small logical steps, with targeted regression tests and app smoke checks after major changes. No publication or remote pushes are assumed.
+- This cloud environment exposes 9 logical CPUs, approximately 9.7 GiB RAM, and no GPU. No connected user computer or saved coding environment is available.
+- Bubblewrap fails namespace setup; PRoot installs but ptrace is prohibited. We will not weaken host security restrictions. Real isolated long simulations may consequently be skipped with evidence.
+- Earlier standalone bundled Gray–Scott reproduction succeeded (four cases, approximately 111 seconds), but does not establish harness or model-driven campaign operation.
+- No live model billing, private credentials, or remote-worker authentication will be used without established authorized access.
+
+## Work plan
+1. Establish broad test and static-analysis baseline; identify reproducible code/API defects and measurable performance opportunities.
+2. Walk first-use GUI workflows, improve shared visual/accessibility patterns and wording, retain existing functionality.
+3. Reconcile docs and contributor workflow with verified rc3 source behavior; draft authorship consideration policy for maintainer review.
+4. Integrate focused fixes, validate existing tests and smoke flows, summarize remaining limitations and decisions.
+
+## Running record
+- 2026-10-01: fetched main and confirmed baseline remains current; created separate audit branch. Existing untracked deployment artifacts are preserved and excluded from commits.
+
+## Final summary — 2026-10-01
+
+Completed the feasible audit and polish pass on `audit/rc3-polish`, based on rc3 commit `2f66b8d`. Source changes were committed in logical steps without merging main, rewriting history, or pushing remotely. This is a review branch, not a release or production-readiness certification.
+
+### Requested work and outcome
+1. **Code/performance:** fixed ledger concurrency and replay identity, extreme-range numerical interpolation, worker polling/staging, resource accounting and failure recovery. Kept refactoring focused; no speculative large-scale rewrite or indiscriminate dead-code removal.
+2. **GUI usability:** fixed wrong-conversation submission/stale-navigation races, draft preservation, retries and disabled-state behavior. Executable JavaScript and local HTTP tests pass; rendered workflow review is blocked.
+3. **Wording:** clarified first-use help, connection errors, SSH expectations, actual backend choices and completion policies.
+4. **GUI consistency:** added shared theme/interface primitives, accessibility semantics, focus treatment, reduced-motion behavior and narrow-layout rules. Final visual review remains necessary on a supported browser.
+5. **Documentation:** reconciled current guides with source/CLI, built 47 public Markdown pages, checked 5,765 local links/assets and parsed 11 Python examples. Historical scientific records were preserved and scope-qualified rather than silently rewritten or claimed freshly reproduced.
+6. **API:** improved nested required/finite-number validation, deadline/timeout validation before persistence, clear client errors and uncertain transport recovery. Valid signatures/schemas remain compatible; invalid inputs and malformed responses now fail more explicitly. Broad API redesign was deferred.
+7. **Contributions:** improved setup/style/PR guidance and added **DRAFT ONLY** paper-authorship consideration wording, with criteria explicitly reserved for Bowen.
+8. **Harness/long simulations:** added regression and synthetic recovery/idle-wait/resource tests. Long isolated WarpX, live provider and remote-worker runs remain untested because prerequisites/authorized access are unavailable. No isolation bypass was used.
+
+### Final verification
+- **Python suite:** 862 passed, 93 failed, 26 skipped (194.49 s). Pristine baseline: 754 passed, 93 failed, 28 skipped. Test-ID comparison found **zero new failing tests**; all 93 baseline failure IDs remain. Added 106 tests, all passing; two previously skipped checks passed after optional dependencies became available.
+- The suite is **not fully green** here. Browser/runtime failures and downstream missing-evidence cascades remain; representative cascades were reproduced and traced to sandbox denial. Rerun the remaining failures on a supported host rather than interpreting this comparison as a full integration pass.
+- **DSH JavaScript suite:** 14 passed, no skips. Optional MCP SDK handshake tests passed. All tracked Python Ruff checks passed; non-vendor frontend JavaScript syntax checks and `git diff --check` passed.
+- **App smoke:** CLI help, workspace/monitor HTTP responses, new static assets and CSP checks passed after changes. Strict Sphinx build and generated-schema consistency passed.
+- **Fresh numerical checks:** built-in kinetic-sufficiency and electrostatic-PIC benchmark commands both passed their expected falsification checks. PIC energy drift: Maxwellian 6.060e-7; two-stream 2.145e-6. These are bounded in-process benchmarks, not WarpX or a live agent campaign.
+- **Packaging:** built the credential-free source archive and checked new CSS inclusion. Fresh-home installer got through download/checksums but was blocked fetching managed Python by TLS `UnknownIssuer`. Certificate verification was not disabled; a complete clean-install smoke remains unverified.
+
+### Measured performance (synthetic, reproducible)
+- 40,000-row / 8-campaign SQLite query fixture: head lookup **1618.25 → 1.35 µs**; incremental replay **294.29 → 49.84 µs**. Index creation/write/storage cost is not included.
+- 1,000 historical-job targeted status poll: **19.884 → 0.072 ms** median. Admission/aggregate status intentionally still reconcile history for correctness.
+- 20 MB provider JSONL fixture: peak Python allocation **40.08 → 0.324 MB**, identical accounting totals; no measured model-token saving claimed.
+- Completed 3 MiB + 7 byte input replay: **4 → 1 staging RPCs**, with worker hash verification retained.
+- Reproduce via `scripts/benchmarks/benchmark_ledger_index.py` and `scripts/benchmarks/harness_polling.py`; timing varies with host load. These are not solver-speedup claims.
+
+### Skipped or limited checks and reasons
+- Full Bubblewrap execution: prohibited network-namespace operation (`NETLINK_ROUTE`). PRoot: prohibited `ptrace`. Neither restriction was weakened.
+- Long WarpX/GPU/SSH campaigns: runtime/isolation prerequisites and verified remote/model access unavailable. No credentials or paid provider calls used.
+- Rendered desktop/mobile GUI and screenshots: Chromium socket creation fails; the separate browser rejects the local URL. DOM-state and HTTP tests do not substitute for visual inspection.
+- Clean installer end-to-end: managed Python download failed certificate verification; source packaging itself passed.
+- Historical scientific/benchmark claims: artifact integrity and documentation consistency checks only, except the explicitly listed fresh numerical benchmarks. External links and every historical numerical claim were not independently revalidated.
+- Repository-wide semantic perfection, security certification and live multi-provider behavior are outside what these bounded checks establish. Detailed per-change outcomes follow.
+
+### Decisions for Bowen
+1. Review GUI appearance, keyboard/mobile behavior and the remaining baseline failures on a Linux host supporting the required execution/browser features before merging.
+2. Decide paper-authorship criteria and discussion/credit process; the draft creates no entitlement or adopted policy.
+3. Decide how CPU reservations should relate to explicit capability thread counts/MPI ranks. This pass fixes per-process CPU-time budgeting but does not silently override pinned OpenMP/BLAS settings, enforce affinity, or impose aggregate MPI quotas.
+4. Review invalid-input/error-contract tightening for downstream clients; decide whether broader API versioning/typing work is desirable. No broad public API redesign was imposed.
+5. Review the historical FLASH interpretation notice and deployment trust boundary for native agents; no historical evidence or agent privilege model was silently rewritten.
+6. Decide when to push/open a PR and which supported host should run long acceptance tests. This branch remains local and unmerged; a review bundle/patch is supplied separately.
+
+## Detailed change record
+
+
+### 2026-10-01 — Atomic event-ledger append
+- Reproduced hash-chain corruption with six concurrent SQLite connections: all writes succeeded but chain verification failed.
+- Reserved the SQLite writer with `BEGIN IMMEDIATE` before reading replay identity/chain head; commit/rollback now encloses the full append. Public method signatures are unchanged.
+- Added concurrent writer/shared-idempotency and failed-insert rollback regression tests.
+- Validation: 99 affected tests passed (ledger, models, parameters, research client, campaign, control, search, proposals, orchestration, web); focused Ruff and CLI help passed. Web tests supply the basic app smoke.
+- No full numerical or autonomous-model claim is made by these tests.
+
+### Baseline validation
+- Pristine rc3 in a separate detached worktree: **754 passed, 93 failed, 28 skipped**, 193.39 seconds. Baseline Ruff across source/tests passed.
+- The unmodified suite is not green in this environment. Many failures require working Bubblewrap or browser binaries; each remaining failure is being compared/classified rather than hidden with blanket skips.
+
+### Indexed campaign replay
+- Added a backwards-compatible `(campaign_id, sequence)` SQLite index; reopening an existing ledger installs it without altering events.
+- Synthetic in-memory benchmark: 40,000 events across 8 campaigns, median of 5 batches of 500 queries. Campaign-head lookup: **1618.25 → 1.35 µs**; incremental-tail retrieval: **294.29 → 49.84 µs**. Query results were checked equal. These are query timings, not end-to-end simulation speedups; indexing incurs storage/write overhead.
+- Validation: 71 affected tests passed; schema export check and Ruff passed. Combined ledger/web/worker/accounting suite: 29 passed; CLI smoke passed.
+
+### Harness polling and usage-log memory
+- Single-job status/artifact polling now reconciles only the selected receipt; aggregate status and resource reservations still reconcile the full job set, but avoid a second history read.
+- Provider accounting streams JSONL and retains accounting fields rather than complete per-request context metadata. Totals/replay precedence remain unchanged.
+- Synthetic 1,000-history-job status benchmark: **19.884 → 0.072 ms** median (276.5×); 20 MB provider-log benchmark: peak Python allocation **40.08 → 0.324 MB** (123.7× reduction), identical totals. This is measured orchestration overhead, not measured model-token savings or solver throughput.
+- Validation: 50 focused tests passed, 14 real-runtime tests skipped; Ruff, CLI smoke and combined web suite passed. Runtime skips are explicitly not successful simulations.
+- Fresh probes confirm Bubblewrap network-namespace and PRoot ptrace operations are prohibited. Long isolated WarpX runs remain skipped; no host security controls were weakened.
+
+### Public MCP input validation
+- Enforce nested required artifact fields (`path` and `sha256`) and reject nonfinite numeric inputs before kernel dispatch. Valid tool schemas and inputs remain compatible; callers relying on malformed inputs will now receive validation errors.
+- Added regressions for missing nested fields, NaN/infinity and valid fractional/integer timeouts.
+- Integration verification: 67 API/client/one-shot/web tests passed, one optional MCP SDK test skipped; generated schemas unchanged and checked; Ruff passed.
+
+### Research client transport errors
+- Convert malformed/empty transport JSON, non-object responses and missing successful results into useful `RuntimeError` diagnostics. Successful response recording and replay behavior are preserved.
+- Added deterministic subprocess-response tests, including successful response persistence. This standardizes invalid-response exception behavior rather than changing valid API signatures.
+- Validation included the 67-test API/web smoke above and focused client tests.
+
+### Canonical idempotency payloads
+- Compare the already-canonical JSON representation for replay identity. Python equality previously conflated `true` with `1` and rejected equivalent tuple/list JSON arrays.
+- Added scalar-representation and equivalent-container tests; clarified the ledger's real single-thread-per-connection contract.
+- Validation: 75 affected tests in focused validation, plus 21 integration ledger/web tests, Ruff and CLI smoke passed. Changed behavior applies to ambiguous or mismatched replay payloads.
+
+### Contribution and security guidance
+- Expanded reproducible bug reports, setup, existing code-style rules, regression/PR expectations, provenance and license guidance.
+- Added **DRAFT ONLY** paper-authorship consideration text. No threshold, guarantee or final criteria adopted; Bowen must decide the process and paper-specific requirements.
+- Corrected security descriptions: native agents are trusted host processes; numerical isolation does not cover them; cooperative PRoot is not an OS security boundary; local UI is not a public multi-user service.
+- Validation: strict Sphinx warning-as-error build and local link checks passed; CLI/schema and no-key artifact replay checks passed. Authorship/security wording still warrants maintainer review.
+
+### rc3 documentation reconciliation
+- Corrected rc2/rc3 installation wording, actual built-in/native backend selection, minimal versus classic deadlines/policies, DSH bundle filenames/tools, SSH worker selection and benchmark-versus-simulation distinctions.
+- Audited 47 public Markdown pages structurally using strict Sphinx/link checks. Current guides received source/CLI spot-check review; historical experiment records received scope/version/link checks, not fresh scientific revalidation.
+- Verified schemas and recorded Gray–Scott/GEM/agent-role artifact verifiers; prepared a no-key coding-benchmark task and exported the leaderboard. Those checks do not run fresh FLASH/WarpX or call models.
+- Preserved historical measured claims and identified uncertain remaining tasks for maintainer review rather than rewriting old evidence.
+
+### Efficient worker input replay
+- Respect the worker's verified complete/size staging acknowledgement, stopping retransmission of already-complete frozen inputs; validate partial offsets and final acknowledgement and reject size changes.
+- A completed 3 MiB + 7 byte replay now takes **1 staging RPC instead of 4**. Worker-side whole-file hash verification is retained; this is a transfer-protocol test, not a remote SSH run.
+- Added partial/completed replay, empty/exact-boundary input and malformed acknowledgement tests. Recovery/web smoke: 26 passed; broader focused harness checks: 60 passed, 14 runtime skips.
+
+### Waiting and benchmark reproducibility
+- Added a synthetic supervisor scheduling test: five pending-job waits cause zero provider turns; a terminal receipt produces one wake. This checks scheduling policy without real model/token billing or numerical execution.
+- Added bounded reproducible ledger and harness microbenchmark scripts in `scripts/benchmarks/`. Fixtures/measurement scope are explicit; runtime variation means repeated timing ratios differ.
+- Validation: synthetic wait test and combined web/recovery smoke passed; scripts passed Ruff and benchmark smoke runs. No actual model-token savings claimed.
+
+### Reject invalid service durations before persistence
+- Reject NaN/infinity, nonpositive values, booleans, strings and unrepresentably large durations before study/experiment state is partially written; failed validation can be retried with valid inputs.
+- Existing valid deadline/timeout signatures are preserved. Added 16 boundary/retry regressions; 87 service-family/web tests passed, plus 27 integration boundary/web tests; Ruff/CLI smoke passed.
+
+### Correct historical summaries and remaining installation drift
+- Qualified FLASH demo summaries: the historical root interval overlaps the allowed band, so the historical ledger label does not independently establish falsification. Original audit JSON, result data and figures are preserved.
+- Clarified legacy launch archive versus workspace installer, legacy monitor DSH default versus conversation backends, and CLI repair versus workspace answer completion policies.
+- Added guidance for reporting blocked/failed/running long checks honestly. Strict docs build passed; 5,765 generated local link/asset references checked without missing targets; 11 Python fenced examples parsed (not executed).
+- Historical scientific measurements, external links, paid models and GPU/SSH solver execution are not newly validated.
+
+### Numerical parameter interpolation
+- Fixed finite extreme log/linear ranges whose intermediate ratio/span overflowed (e.g. log midpoint of [1e-300,1e300] incorrectly mapped to upper bound rather than 1).
+- Retained ordinary seeded-design formulas exactly; only overflowing intermediate cases use weighted interpolation. Added six extreme-range regressions.
+- Validation: 58 parameters/models/search/campaign/domain/web tests passed; parameter/GUI/web integration: 46 passed; Ruff and CLI smoke passed.
+
+### GUI workflows, wording and shared interface primitives
+Problems found and fixed:
+- A conversation change while submitting could send a message to the wrong conversation, clear a newer draft or apply stale project data. Capture destination/draft and reject stale navigation results.
+- Generic action cleanup re-enabled Send despite the backend rejecting messages during an active agent run. Preserve that disabled state; repeated benchmark submits are guarded.
+- Failed API connection saves left misleading pending text; now report failure and allow retry. SSH password input clears when its dialog closes.
+- Workspace and monitor had divergent theme behavior; share preference bootstrap and common typography, palette/focus/spacing/radius primitives. Added reduced-motion treatment, constrained dialogs and narrow benchmark layout.
+- Added keyboard skip navigation without corrupting hash routes, named dialogs, current-navigation/mode accessibility semantics and concrete first-use/SSH instructions.
+Validation: 72 GUI/workspace tests passed, one pre-existing namespace-dependent skip; integration 46 passed. All non-vendor JavaScript syntax checks, Ruff and local HTTP/CSP asset smoke checks passed.
+**Limit:** actual rendered desktop/mobile/browser QA and screenshots remain blocked. Shell Chromium cannot create required sockets, and the separate cloud browser rejects the local URL. No rendering or visual-polish certification is claimed; Bowen should review the appearance on a supported browser.
+
+### Allocated CPU-time budget and uncertain worker replies
+- Account for explicitly allocated CPUs in per-process `RLIMIT_CPU` seconds so threaded jobs do not lose most of their wall-clock allowance. None/one-CPU behavior, other limits and explicit capability OpenMP/BLAS settings remain unchanged. This is not CPU affinity, an aggregate MPI quota or a host security-policy change.
+- Example constructed limit: 8 allocated CPUs × 2.5 seconds yields 21 CPU seconds including cushion, versus the old 4. Invalid/nonfinite allocations are rejected before limit calls. Tests mock limit calls rather than changing host limits.
+- Treat malformed/inconsistent worker success envelopes as transport uncertainty (`WorkerUnavailable`), retaining genuine worker rejections as semantic errors; do not prematurely fail a possibly running experiment.
+- Validation: CPU construction/transport/staging/web suite passed, including a recovery test for malformed replies; CLI and Ruff passed. Actual multi-core isolated runs remain untested due host restrictions.
+
+### Independent review corrections
+- Independent review found that targeted polling could leave unrelated stale jobs counted against admission. Reconcile once under the existing admission lock; per-job polls stay O(1). Four saturation regressions reclaim 256 expired/disappeared/cancelled receipts and preserve the genuinely-live queue cap.
+- Independent review also found that generic bootstrap transport legitimately returns scalar success values. Keep generic result compatibility while requiring object results only at the worker-RPC boundary.
+- These audit-introduced integration regressions were fixed in follow-up commits before final validation, without rewriting history.
+- Validation: independent harness review 49 passed; focused harness scope 88 passed, 14 environment-gated skips; review/web regression smoke passed. Two hundred seeded randomized provider logs matched pristine accounting totals. Ruff passed.

{simjecture-0.5.2rc2 → simjecture-0.5.3}/CHANGELOG.md

@@ -2,6 +2,208 @@
 
 This project follows semantic versioning. Dates use ISO 8601.
 
+## Unreleased
+
+## 0.5.3 — 2026-10-02
+
+Stable release of the 0.5.3 candidate series. Minimal remains the default;
+structured and frontier remain selectable. Existing research data and older
+program folders are preserved by the versioned installer.
+
+- Run recorded numerical experiments on local or SSH workers with CPU/RAM/GPU
+  reservations, automatic machine setup, 30-second availability checks and recovery
+  after interruptions. Research agents and independent review stay on the coordinator.
+- Publish the measured coding-agent leaderboard with task-specific time/cost
+  rankings, failure-inclusive cost bars, consistent model colours and a Pareto plot.
+  Custom trials and community grades remain separate; the benchmark analyzes
+  recorded diagnostics and does not launch new simulations.
+- Repair concurrent ledger writes, idempotent replay, reserved-CPU time limits,
+  transport recovery, input validation and conversation races. Retain request-level
+  usage records, native DeepSeek tool-message continuity and review evidence retrieval.
+  Preserve scientific review gates, deadlines and the benchmark task pack 0.3.0.
+- Connect conversation, selected study, and **Evidence & review** with shared
+  navigation. Study links retain their selection across reloads, mode changes,
+  and monitor Back/Forward navigation.
+- Return from study details to the owning conversation, and prepare continuations
+  there by default. Standalone CLI studies and read-only inspection remain available.
+- Guard study rendering and continuation dialogs against stale requests, newer
+  navigation, and repeated submission. Clarify that simulation output alone is
+  not an independently accepted scientific finding.
+- Release pending continuation previews when navigating away, so a stalled old
+  request cannot block a new action. Ignore its late reply even after returning
+  to the original route; retain the currently open dialog.
+- Use shared light/dark logo artwork in the workspace and monitor, with SVG/ICO
+  favicons, accessible home links and packaged font attribution. Keep assets
+  self-contained and preserve the existing content-security policy.
+- Make the README and documentation entry points workspace-first, move terminal
+  automation into a dedicated guide, and group dated reports in an archive index.
+  Simplify contributor setup and checks, and keep release facts in the changelog
+  and acceptance records.
+
+## 0.5.3rc4 (preview) — 2026-10-02
+
+- Integrate the independent rc3 audit with its reproducible fixes and regression
+  tests. Acquire the SQLite write transaction before reading the ledger head;
+  concurrent writers now preserve the event hash chain. Compare canonical JSON
+  for idempotent replay and index campaign event heads and incremental replay.
+- Scale per-process CPU-time allowances by explicitly reserved CPUs, preventing
+  threaded experiments from being killed before their wall-time allowance expires.
+  This does not impose CPU affinity or an aggregate MPI process-tree quota.
+- Reconcile only the requested worker receipt during targeted polling, while
+  preserving full reconciliation for admission and reservations. Avoid sending
+  remaining input chunks after a checksum-verified completed-upload acknowledgement.
+- Preserve uncertain execution state after malformed worker replies, distinguish
+  worker RPC object results from valid scalar bootstrap results, and improve
+  research-client transport errors.
+- Stream provider usage logs while retaining request accounting fields. Reject
+  non-finite numbers, missing nested required inputs and invalid durations before
+  persistent side effects. Avoid overflow when mapping finite extreme parameters.
+- Fix stale conversation navigation, wrong-conversation submissions, draft loss,
+  duplicate benchmark submission and retry/button state. Clear SSH passwords when
+  closing the editor and report API connection save failures inline.
+- Add shared workspace/monitor styling, keyboard skip links, accessible labels,
+  reduced-motion support and bounded scrolling dialogs. Main page layouts are
+  retained; this is a modest interface polish pass.
+- Reconcile installation, permissions, native-agent and research documentation
+  with current behavior. Qualify historical reconnection conclusions against their
+  registered tolerance band without changing original scientific records.
+- Keep minimal as the default and retain existing research strategy, review gates,
+  completion rules, benchmark task pack 0.3.0 and recorded model measurements.
+  No new live-model or long scientific campaign is claimed by this release.
+- Explicitly exclude installed dependencies and local environments/state from
+  source distributions, including when building from a Git worktree.
+
+## 0.5.3rc3 — 2026-10-01
+
+- Add separate finish-time and cost ranked bar plots for every published task.
+  Show unfinished configurations in red with recorded cost bars; omit no-inference
+  configurations from public rankings;
+  plot comparable completed profiles on an upper-left-preferred cost–time Pareto
+  chart with an observed frontier and model-effort connections.
+
+- Present a maintained official benchmark edition by default, with two task tabs,
+  explicit time and cost ranks, input/output API tariffs, model search and effort
+  filters. Keep prior qualifications in audit records and current contributed/local trials
+  in their own comparison groups.
+- Plot API-equivalent cost versus task time with dashed lines connecting model
+  effort settings; preserve partial receipt bounds and labelled AGY estimates.
+- Recover a separate valuation ledger from saved usage counters without changing
+  host grades, and export a standalone read-only leaderboard page without model calls.
+
+- Add a task-scoped benchmark leaderboard with pass-rate uncertainty, separate
+  actual reported spend and API token estimates, failure-inclusive cost/time per
+  success, and descriptive Pareto comparisons after repeated trials.
+- Redesign Benchmarks with task cards, comparison groups, time/cost rankings,
+  a cost/time plot, graded-trial import, private-field-free summary export,
+  dated pricing references and a linked methodology. Remove pilot/unranked panels
+  and the completion/time view from the public page. Model colours are distinct;
+  outlined frontier points replace connections between different models.
+- Preserve trial identity during regrading and report the agent that actually
+  performed benchmark turns, with request-level usage where available.
+- Expose `max` reasoning effort in native launch contracts and the workspace
+  selector, allowing supported models to be compared at each effort setting.
+- Add fresh timed native/API model trials, arbitrary custom model IDs, a visible
+  Run your model dialog, campaign progress, and portable sanitized grade bundles.
+- Repair equal-radius tracer interpolation in independently versioned task pack
+  0.2.0 and float32 intermediate rounding in 0.3.0; preserve affected qualification
+  grades as unranked records and verify precision with independent scalar integrals.
+- Preserve cumulative AGY usage without double counting and partial native Grok
+  receipts; expose native `ultra` reasoning effort where supported.
+- Ship actual host grade bundles and a community submission directory, keeping
+  contributor-declared measurements in separate comparison groups.
+- Complete an owned 42-configuration, 252-attempt sweep with five CSV attempts
+  and one recorded RZ diagnostic attempt per configuration; publish the full
+  result table and audited receipts. Separate no-inference quota/expiry from
+  interrupted paid work, and retain usage preceding zero-counter AGY errors.
+- Accept downloaded grade bundles in the CLI leaderboard command and preserve
+  a selected virtualenv Python path for subsequent timed trials.
+
+## 0.5.3rc2 (preview) — 2026-09-30
+
+- Polish Research tools with compact independent cards and a separate details
+  panel. Keep paths, diagnostics and logs accessible without expanding grid rows;
+  distinguish detected local builds from unavailable managed-runtime profiles.
+- Add standalone local/SSH headless numerical workers for minimal studies, with
+  frozen runtime identities, CPU/RAM/GPU admission and asynchronous execution.
+- Preserve job identity across lost replies, SSH interruptions and coordinator
+  restarts; enforce worker deadlines and confirm descendant cancellation.
+- Add the Machines workspace page, visible study pool selection, placement and
+  transport status, bounded read-only source access and verified artifact retrieval.
+- Simplify SSH onboarding to address and password with automatic hardware/runtime
+  detection, optional advanced settings, recorded agent setup commands and independent
+  30-second availability heartbeats. Polish machine cards, jobs and mobile navigation.
+- Install the matching released package on headless workers instead of a fixed
+  older candidate.
+- Preserve pool selection when preparing continuation phases. Keep research agents
+  and independent scientific review central; Simote remains optional.
+
+
+## 0.5.3rc1 (preview) — 2026-09-30
+
+First repairs from the aluminium stagnation campaign audit. Minimal remains the
+default; this preview does not change worker session policy or claim scientific
+improvements from a new long campaign.
+
+- Recognize provider `402 Insufficient Balance` errors in native result envelopes.
+  Pause with a credit/quota explanation and keep Resume and the original deadline.
+- Explain and evaluate alternative instrument requirements as **any of**, in methods
+  review, claim review, study status and the browser; FLASH or WarpX never means both.
+- Bound long plan notes before evicting current evidence. Preserve references to the
+  newest completed experiment, method and claim review even in small recovery briefs.
+- Let methods/progress reviewers request recorded JSON excerpts through the host,
+  with hash checks, bounded size and at most two retrieval rounds. Requests and failed
+  retrievals cannot grant approval; record each review packet and response.
+- Record builtin provider requests before agent-step aggregation, including completion
+  retries, failures, native cache/reasoning usage when reported, and context sizes/hashes.
+  Preserve unknown counters and avoid adding request and legacy step totals together.
+- Show live request counts, worker/reviewer token breakdown, missing usage, provider
+  retry time and actionable provider errors in terminal/browser views. Monetary cost
+  is explicitly unavailable without billing data; no price assumptions are invented.
+- Preserve native tool-call/result messages in builtin DeepSeek worker conversations,
+  including private reasoning continuity. Nineteen controlled live repair trials
+  informed this choice; retain the completion guard, session policy and review gates.
+  Record response finish reasons, call counts and text trimming for further audits.
+- Ship experimental **Simjecture Bench 0.1.0**, independently versioned CSV and
+  recorded-RZ diagnostic coding tasks, portable CLI preparation/grading and Harbor
+  task-format export. Verify varied numerical holdouts, saved outputs, input hashes
+  and numerical plot data; keep scientific claim approval separate.
+- Add a visible Benchmarks workspace entry for preparation with the chosen agent
+  and independent grading after delivery. Preparation makes no provider request.
+  Timed framework trials use the external runner; interactive chat is exploratory.
+- Document the matched MiMo 2.6 Flash/Pro and DeepSeek V4.1 Flash pilot, including
+  cache usage, continuation/setup differences and limits of single-trial evidence.
+
+See [audit and next phases](docs/research/stagnation-deep-audit-20260929.md) and
+[preview scope](docs/testing/0.5.3rc1-acceptance.md), and
+[adapter benchmark](docs/testing/native-api-adapter.md).
+
+## 0.5.2 — 2026-09-29
+
+Stable release of the conversation-first research workspace tested in rc1/rc2,
+plus continuation, steering, review recovery and the ITER diagnostics pack.
+Minimal mode remains the default; structured/frontier modes remain selectable.
+The versioned installer preserves existing research data and older program folders.
+
+- Add an optional ITER diagnostics/data pack with pinned CHERAB/Raysect,
+  CHERAB-ITER/IMAS extensions and IMAS tools. Install/check from the CLI or GUI;
+  run a visible, monitored numerical demo with plots. Add source-based agent
+  guidance and separate setup entries for JOREK, SOLPS-ITER and DINA-PS.
+
+- Offer Prepare directly or Prepare with agent from the continuation dialog.
+  Chat preparation carries parent context across turns and lets the agent draft
+  the brief and select inherited files without losing lineage or instruments.
+- Add Continue investigation and Send guidance to study cards and the experiment
+  monitor. Review a linked phase's brief, selected inherited files, model and new
+  budget before launch; retain the original study and evidence status.
+- Snapshot continuation context with verified hashes and immutable parent lineage.
+  Add CLI continuation and advisory steering; show queued/checkpoint delivery in
+  the web interface and include guidance in independent review packets.
+- Detect empty/truncated built-in review replies, record finish reasons and usage,
+  and use bounded recovery with a larger output allocation and a DeepSeek
+  non-thinking retry. Back off repeated invalid oversight without granting approval.
+- Reconcile expired deadlines/dead supervisors in live status and write final
+  reports when a supervisor exits through an external error.
+
 ## 0.5.2rc2 (preview) — 2026-09-27
 
 - Provision scientific compilers, MPI, HDF5 and other prerequisites in managed

{simjecture-0.5.2rc2 → simjecture-0.5.3}/CITATION.cff

@@ -2,8 +2,8 @@ cff-version: 1.2.0
 message: "If you use this software, please cite the exact software version and Git commit."
 title: "Simjecture"
 type: software
-version: 0.5.2rc2
-date-released: 2026-09-27
+version: 0.5.3
+date-released: 2026-10-02
 abstract: >-
   An evidence-governed runtime for autonomous computational experimentation and
   falsification over hypothesis trees.

simjecture-0.5.3/CONTRIBUTING.md

@@ -0,0 +1,105 @@
+# Contributing
+
+Thanks for helping improve Simjecture. You can contribute without running a full
+research campaign or installing a scientific solver. Useful contributions include:
+
+- reporting a reproducible bug or a confusing part of the interface;
+- improving an example, installation step, or explanation;
+- adding a small deterministic regression test;
+- improving a runtime integration, scientific validation, or evidence workflow.
+
+For a first change, pick one unclear documentation step or an edge case in an
+existing test module. Keep it small enough to explain and review on its own.
+Discuss large changes or scientific-policy changes with the maintainer before
+investing in an implementation.
+
+## Report a problem
+
+Include the exact package version and Git commit, operating system, execution
+backend, minimal reproduction, expected result, and observed result. Identify
+whether you are using a stable release or a preview; their features may differ.
+Redact credentials and private paths. For security-sensitive issues, follow
+[SECURITY.md](SECURITY.md) rather than posting the details publicly.
+
+## Start developing
+
+Use Linux (or WSL), Python 3.11+, and uv. From a source checkout:
+
+```bash
+uv sync --frozen
+uv run pytest tests/test_models.py
+```
+
+The default `dev` dependency group includes pytest and Ruff. The example test
+module checks model validation without provider credentials or a solver. Install
+additional dependencies only when they are needed for your contribution.
+
+## Check the part you changed
+
+- **Python:** run `uv run ruff check .` and the affected test modules, then broaden
+  coverage as appropriate. Follow nearby code, use clear names and consistent type
+  annotations, and avoid unrelated reformatting. Ruff's rules are in
+  [pyproject.toml](pyproject.toml)
+- **Public models:** regenerate with `uv run simjecture schemas --output schemas`,
+  then run `uv run simjecture schemas --output schemas --check` and relevant tests
+- **Documentation:** use `uv sync --frozen --group docs`, then
+  `uv run --group docs sphinx-build -W --keep-going -b html docs docs/_build/html`.
+  Check changed examples and links; see [documentation development](docs/development/documentation.md)
+- **Browser UI:** add the `browser` dependency group with
+  `uv sync --frozen --group browser`, install Chromium with
+  `uv run playwright install chromium`, and run the affected browser tests.
+  Add `--extra workspace` to the sync command when testing workspace features
+- **DSH integration:** use `uv sync --frozen --extra dsh` and the Node version
+  pinned in [CI](.github/workflows/ci.yml), then run `npm ci --prefix integrations/dsh`
+  and `SIMJECTURE_MCP_EXECUTABLE="$PWD/.venv/bin/simjecture-mcp" npm test --prefix integrations/dsh`.
+  See the [adapter guide](integrations/dsh/README.md#adapter-validation)
+- **Execution backends and scientific runtimes:** follow the relevant capability
+  guide and probe the actual backend before testing. See
+  [adding a capability](docs/how-to/add-a-capability.md) and
+  [restricted hosts](docs/how-to/restricted-containers.md)
+
+The [CI workflow](.github/workflows/ci.yml) defines the broader integration and
+release checks; its full environment is not the starting requirement for every
+contribution. Bubblewrap needs working user namespaces, and cooperative PRoot
+needs working tracing. Record unavailable prerequisites, failures, and skipped
+checks separately from passes, and run the independent checks you can. Do not use
+real provider credentials or paid model calls for ordinary regression tests.
+
+## Keep research records trustworthy
+
+Simjecture is a research harness: agents choose scientific details; the harness
+controls permissions, provenance, resource limits, and evidence eligibility.
+Keep UI projections separate from durable scientific state, validate untrusted
+inputs at their boundary, and do not bypass permission or evidence gates.
+
+Label fixtures, commissioning, discovery, and held-out confirmation clearly.
+Preserve original records and append corrections. A replay or coding benchmark
+on recorded diagnostics is not a fresh simulation or scientific acceptance.
+Keep conclusions within their evidence contract and numerical qualification.
+For the workflow-specific rules, read [evidence and claims](docs/concepts/evidence-and-claims.md)
+and [minimal-mode methods review](docs/how-to/minimal-oversight.md).
+
+## Before opening a pull request
+
+- Use a focused branch from the intended base, and explain the change and any
+  compatibility or migration effects
+- Add deterministic regression tests for changed behavior, including relevant
+  failure, restart, and recovery paths. Use fixtures and temporary directories
+- Update affected examples, user documentation, and the changelog for user-visible
+  changes. For UI changes, include desktop/narrow screenshots and the interaction tested
+- List the commands you ran and their actual outcomes, including blocked or skipped
+  checks. Keep local results separate from historical release evidence
+- Review the diff for secrets, generated output, and unrelated changes, then request
+  maintainer review. Scientific-policy, security-boundary, and release changes
+  require appropriate maintainer review before merging
+
+## License and provenance
+
+By contributing, you agree that your contribution is licensed under the
+[Apache License 2.0](LICENSE). Identify upstream sources and retain required
+copyright/license notices. Do not contribute code, data, solver material, or
+unpublished third-party artifacts you lack permission to redistribute. Keep local
+runtime installations, bulk simulation output, and credentials out of commits.
+Cite scientific and software sources precisely, including versions and commits
+when reporting results. See [CITATION.cff](CITATION.cff) and
+[third-party notices](THIRD_PARTY_NOTICES.md).

simjecture-0.5.3/NAVIGATION_INTEGRATION.md

@@ -0,0 +1,108 @@
+# Shared study navigation integration log
+
+## Scope and base
+
+- Work date: 2026-10-02 (UTC)
+- Local branch: `feat/rc4-shared-study-navigation`
+- Exact integration base: `cc4ae638cc69911f2fd252cb1bbe8bee4cabf149`
+  (released `v0.5.3rc4`, also current upstream main when checked)
+- rc4 includes the exact audit head `c13585139973d7ab2d5057938db2f3c332b5c64d`
+  above public rc3 `2f66b8d06c20b18cfea20188acad60b83fa25ec0`.
+- The first read-only upstream check still found rc3. After the user's rc4 update,
+  a fresh check verified annotated tag `v0.5.3rc4` and its peeled commit above.
+- The original audit-based navigation branch was preserved. A new rc4 worktree
+  received only the seven incremental navigation commits. All implementation/test
+  commits applied cleanly because rc4's web code matches the audit head. The sole
+  conflict was the changelog: rc4's release section was retained and the new
+  navigation notes remain under Unreleased. No audit commits were duplicated.
+- rc4 version, installer, packaging exclusions, release acceptance notes, and
+  dependency-lock changes are preserved. No push, merge, release, deployment,
+  history rewrite, credential use, or paid model call was performed.
+
+## Implemented
+
+1. Shared URL builders and a semantic Conversation → Study → Evidence & review
+   trail in the existing workspace and study-details pages.
+2. Explicit selected-study cards and study-preserving conversation/mode links.
+   Reloads and newly launched studies retain an unambiguous selection.
+3. Study details resolves its owning conversation from server-side, path-checked
+   workspace records. The return and continuation links use that exact owner.
+4. Implicit continuation reuses an owned conversation, with busy and pending-brief
+   guards. Direct CLI and read-only inspection remain available; standalone
+   continuation retains its create-on-submit fallback.
+5. Monitor campaign choices update the URL and respond to Back/Forward. Stale
+   snapshot/study-list responses cannot overwrite a newer choice.
+6. Continuation dialogs capture their source context, suppress duplicate requests,
+   preserve retryable form errors, and dismiss on route departure. Late results
+   do not drag a user away from newer navigation. Before submission, Cancel/Escape
+   does not prepare a proposal. During a pending POST, Cancel/Escape is disabled
+   because cancellation could not guarantee the server had not prepared it.
+7. Ordinary simulation results are explicitly distinguished from independent
+   scientific acceptance. Graphs, evidence, advanced controls, and scientific
+   storage/contracts are retained; this is not a frontend rewrite.
+
+## Validation record
+
+See the review pack's `TEST_RESULTS.md` for final commands, counts, failure
+comparison, and bundle/patch checks. New coverage includes ownership, standalone
+fallback, legacy launched briefs, read-only state, captured continuation identity,
+repeated/cancelled flows, reload/mode state, navigation races, and monitor history.
+
+During integration, the local HTTP/CSP test caught the missing server allowlist
+entry for the new JavaScript module. It was corrected in a separate commit and
+retested. A successful syntax check alone would not have caught that issue.
+
+## Known limits and review needs
+
+- Rendered desktop/mobile browser QA and screenshots were unavailable. The
+  Playwright-managed Chromium binary is absent; the prior system Chromium attempt
+  could not create required sockets, and the cloud browser rejected local URLs.
+  No security restriction was bypassed. Deterministic DOM tests and HTTP checks
+  are not a visual inspection. Runnable browser acceptance tests are included.
+- The full inherited suite has known environment/runtime failures; the review
+  pack distinguishes current outcomes from historical audit results.
+- Standalone continuation has no durable request-key replay journal. After an
+  ambiguous lost POST response, check for the created `Continue:` conversation
+  before resubmitting. Repeated pending clicks are suppressed; owned retries
+  reuse the owner and protect an existing proposal.
+- Navigation selection is URL-backed, not a server-side last-view preference.
+  Opening a conversation without a study in its link selects its latest study.
+- Existing unrelated unlaunched proposals are not silently cleared. Resolve the
+  proposal or active conversation turn before preparing another continuation.
+
+## rc4 integration guidance
+
+This pack is incremental from the exact released rc4 commit named above. It
+contains only the navigation integration, not the earlier audit fixes or rc4's
+release/packaging changes. Import the bundle or apply the patch on a separate
+branch starting at that rc4 commit. Both routes were checked to reconstruct the
+same final Git tree.
+
+For a later branch that has changed these areas, likely overlap points are
+`web/workspace.py` (continuation/ownership), `web/application.py` (snapshot
+projection), `web/server.py` (asset allowlist), `web/static/workspace.js`
+(routes/actions/study cards), `web/static/app.js` (monitor selection/control
+links), both HTML shells, and `interface.css`. Preserve later features and the
+new regression tests when resolving conflicts. Keep conversation ownership
+authoritative and path-checked, and do not relax read-only controls or independent
+evidence semantics. Rerun validation after conflict resolution; the current
+results apply to the supplied rc4-based tree.
+
+## Local audit and repair — 2026-10-02
+
+The subsequent local audit verified the kit on the exact released rc4 base,
+ran all six rendered browser cases, and completed a supervised full suite with
+1,041 passes and seven optional-runtime skips. Desktop/mobile checks covered
+light/dark study, evidence and continuation views.
+
+A delayed-response check found that leaving before a continuation preview
+created its dialog retained the global action lock. New actions stayed blocked
+until the old request returned. The repair clears the departed token regardless
+of whether a dialog exists and checks token identity before opening a preview.
+Regression coverage exercises a new action while the original request remains
+pending, plus returning to the same route before its reply. The two deterministic
+cases failed on the supplied implementation before the repair.
+
+This is a local development preview; the published rc4 tag and artifacts retain
+their original contents. No new model inference or scientific campaign is needed
+to inspect the navigation.