OpenPuzzle 1.0.53 is available.

Previous release · 30 October 2026

OpenPuzzle 1.0.52

Vanity CUDA compatibility: The Vanity benchmark now launches its native engine with the correct process owner. Native errors are visible if initialization fails. Automatic GPU tuning, the one-hour range target and common controls are preserved.

Architectures: native sm_50, 52, 60, 61, 70, 75, 80, 86, 89 and 90, plus compute_50 PTX. Driver support is required; compiling a target does not validate every GPU model.

Validation: earlier six real known-address results preserved; three affected synthetic/UI suites and packaged-engine SHA-256 checks. The engine itself is unchanged. Private keys remain local.

Also available: Linux TGZ · Versioned source archive. Finish active assignments with Safe Stop and close the application before updating.

Ubuntu package SHA-256:
4b4942483683f44379ba1de60354363c777d6713114025271103df4983b48433

Previous release · 30 October 2026

OpenPuzzle 1.0.51

Vanity CUDA compatibility: Vanity CUDA now includes multiple NVIDIA architectures. Known-address tests passed on RTX 4070 SUPER and all five RTX 3080 cards individually. Automatic GPU tuning, the one-hour range target and common controls are preserved.

Architectures: native sm_50, 52, 60, 61, 70, 75, 80, 86, 89 and 90, plus compute_50 PTX. Driver support is required; compiling a target does not validate every GPU model.

Validation: six existing real known-address tests rechecked; three affected synthetic/UI suites, independent CPU references and packaged-engine SHA-256 checks. Private keys remain local.

Also available: Linux TGZ · Versioned source archive. Finish active assignments with Safe Stop and close the application before updating.

Ubuntu package SHA-256:
cfd7df7b047dd10303a36c52a1a87c42c2280a88b578cad11c9de0af990ed092

Previous release · 29 October 2026

OpenPuzzle 1.0.50

Common Stop behaviour: Stop cancels the Vanity assignment like BitCrack, allowing the server to recycle the range. Safe Stop finishes the current assignment. Cancellation retries preserve the original request and do not relaunch the motor.

Nodes: the ordinary heartbeat now reports Vanity / CUDA and detected installation state to the existing private Nodes page. No API or database change is required.

Preserved behaviour: recovery after interruption, found-key export, benchmarks, GPU profiles, geometry and the usual one-hour range target remain in place.

Validation: 11 affected suites passed, including synthetic single/multi-GPU Stop and Safe Stop, HTTP cancellation retries and heartbeat inventory. Earlier real RTX 4070 SUPER known-address evidence was preserved. This correction was prepared without a real GPU run or new benchmark; real rig acceptance and production solution notices remain unverified.

Also available: Linux TGZ · Versioned source archive. Finish active assignments with Safe Stop and close the application before updating.

Ubuntu package SHA-256:
695ee461da4f46b9421b12361d0f86307f7633fadf77e9ddc9b7ccf0ac7f67e8

Previous release · 29 October 2026

OpenPuzzle 1.0.49

Vanity CUDA: integrated into the normal client beside BitCrack, KeyHunt and Kangaroo, using the common Start, Stop, Resume, Safe Stop, status, progress and private solution export. The desktop header displays the version.

Measured profiles: an automatic benchmark selects the geometry for each GPU before its first Vanity assignment. Assignment sizes follow measured speed and the usual target of about one hour. Consecutive native blocks stay within the same assignment; Safe Stop finishes it before blocking new work.

CUDA rigs: use openpuzzle run 71 --engine vanity --backend cuda --devices all. The existing supervisor provides one slot and assignment per physical GPU.

Validation: affected automated suites passed. A real RTX 4070 SUPER found a known test address; the exact key, private wallet export, engine termination and local HTTP notification were verified. Multi-GPU supervision passed synthetic tests. Real multi-GPU rig acceptance and a production website solution notice have not been verified. Earlier full-suite results were preserved without rerunning them.

Availability: Linux x86_64. Vanity supports CUDA and linear P2PKH puzzles 65–80. Windows remains a separate experimental release. Private keys are never uploaded; solution reports contain only client and assignment identifiers.

Also available: Linux TGZ · Versioned source archive. Finish active assignments with Safe Stop and close the application before updating.

Ubuntu package SHA-256:
99157fe96e6c112fe91499d9ef3266bbc69a6bec596a7456637d8f46e2dcfee5

Previous release · 21 October 2026

OpenPuzzle 1.0.48

Automatic desktop benchmarks: choose Benchmark in the Linux panel, select CUDA or OpenCL and the device, then confirm the local measurement. Parameters are selected automatically. Progress appears in the panel and a successful benchmark saves the measured GPU profile.

Cancellation and preserved profiles: Stop cancels the benchmark. A failed or cancelled measurement preserves the previous profile. Stop active searches before measuring. The benchmark does not change GPU power, clocks or fans.

Measured tuning: the shared candidate matrix, stability checks, 60-second finalist and 3% planning margin are included. GPU power and fan controls from 1.0.47 remain available through the authenticated helper.

Validation: 178 native Linux suites and 24 isolated platform suites passed on the versioned source. The owner separately confirmed the real Ubuntu benchmark and the installed interface. Existing engine binaries were reused unchanged.

Platform scope and privacy: Linux 1.0.48 is stable. Windows Setup 0.1.5 remains a separate experimental release. Private keys remain local; the public website and network-status API expose no telemetry or physical GPU identifiers.

Also available: Linux TGZ · Versioned source archive. Finish active assignments with Safe Stop and close the application before updating.

Ubuntu package SHA-256:
be83326a6cf855a0de9f74529290c8ca8b83a287cf2d2cbeafa4a9fda3d8ef2a

Previous release · 17 October 2026

OpenPuzzle 1.0.47

GPU power and fans: the Linux desktop panel now includes power limits and Auto / Manual fan controls for supported NVIDIA and AMD boards. NVIDIA uses NVML; AMD uses amdgpu. Auto returns fan control to the driver; Manual sets a fixed percentage. Support varies by board and driver.

Apply to the selected GPU: the panel verifies physical GPU identity before each change, validates supported ranges and reads back the result. Failed changes attempt rollback and report any restoration failure. Opening or refreshing the panel does not change GPU settings. Manual fan settings require at least 30% and a valid GPU temperature below 80 °C when applied. Existing thermal search protection remains separate.

Authenticated controls: Linux packages include a dedicated root-owned helper and restricted polkit action. Apply requires administrator authentication while the desktop panel stays unprivileged. A graphical polkit agent and pkexec must be available. Installation does not apply tuning or start a search. Confirmation dialogs are readable in light and dark themes.

Validation: all 176 automated client tests passed on the versioned Linux source, including GPU transaction, desktop confirmation and package-content checks. NVIDIA and AMD controls were also confirmed manually on Linux. Existing engines and measured profiles are preserved.

Platform scope and privacy: this stable release ships Linux packages. The separate Windows 11 x64 public test remains experimental. Private keys remain local. The public website and network-status API expose no telemetry or physical GPU identifiers.

Also available: Linux TGZ · Versioned source archive. Finish active assignments with Safe Stop before replacing the installed client.

Ubuntu package SHA-256:
21d29a8bf30dcc5d8bd1c5707bcaa7b2e9f6908202ac2d0c39c99e244010f1e9

Previous release

OpenPuzzle 1.0.46

Complete integer arguments: affected CLI commands reject missing or empty values, fractions, exponent syntax, whitespace, trailing text and overflow. Defaults apply only when an option is absent. Signed decimal values and leading zeroes remain accepted within the integer range, subject to the bounds of each command. Duplicate numeric options and aliases fail instead of silently choosing a conflicting value.

GPU selection and benchmark: invalid numeric input is checked before saved settings, device initialization or Rusticl updates are changed. Benchmark numeric syntax errors report OP-BENCH-011. Existing benchmark duration, sample and launch-parameter bounds remain in place.

Validate before job and runtime startup: job commands check numeric options before database initialization. Run commands validate puzzle, GPU and OpenCL device indices, launch dimensions, CPU threads and assignment duration before runtime acquisition, preserved execution recovery or assignment requests. Multi-GPU and concurrent supervisors use the same early checks. An invalid numeric request cannot trigger these operations through dry-run.

Validation: all 154 automated client tests passed before packaging. The grouped corrections were made in one client commit. CliIntegerTests and CliArgumentIntegrityTests cover numeric boundaries, missing values, duplicate aliases and rejected requests that preserve local configuration and execution state. Tests use isolated fixtures without registration, assignment requests or real engine searches.

Platform availability: this release provides Linux packages. A validated Windows executable is not yet included. Search engines and the deployed cancelled-range recycler remain unchanged.

Privacy boundary: GPU readings remain in authenticated client heartbeats and the private Admin Nodes dashboard. The public website and network-status API expose no telemetry or physical GPU identifiers.

Package SHA-256: 873fb9206907769ac76fdec1dd43b77cdbd0a24d048fa545b9f0a4b2e9e44aa2

OpenPuzzle 1.0.45

Preserve settings when a control update cannot read them: thermal CLI updates, desktop thermal saves, GPU selection and benchmark selector updates preserve existing empty, malformed or unreadable configuration files instead of replacing them with defaults. Only a missing configuration supplies creation defaults. The controls report failure; benchmark stops before context/device initialization with OP-BENCH-010.

Configuration documents are objects: arrays and scalar roots are rejected rather than reading unrelated entries as settings. Compatible flat and nested object layouts and established typed-field defaults remain supported. Read-only consumers retain their existing fallback defaults.

Read without blocking or changing a symlink target: nonblocking configuration reads reject non-regular files and final-path symlinks. FIFOs cannot hang the reader, and symlink targets are neither read nor chmodded. Owner-only permissions are applied to the opened regular-file descriptor instead of following a changed path.

GPU selection retains supported settings: choosing a device preserves supported engine/backend/executable paths, thermal policy, Rusticl selector and assignment duration through the existing atomic writer. Negative device indices fail. The CLI returns failure instead of printing success when loading or saving fails. Saved-device reads share strict parsing, so fractional numbers cannot select a numeric prefix as a GPU identifier.

Validate benchmark arguments before changing Rusticl: invalid backend, device, duration, sample count or launch parameters leave configuration bytes and the process Rusticl environment unchanged. Validation happens before applying or persisting the selector. Regression tests use invalid requests before GPU discovery and do not run real benchmarks.

Limits and recovery: damaged settings are not repaired automatically. Preserve the original file and recover known settings from a backup before retrying. Unknown configuration fields retain the existing full-save behavior and are not preserved by this change. Parent-directory symlinks and concurrent read/modify/write cycles are not covered. Explicit full saves remain available; atomic publication does not add a transaction across files or guarantee directory-entry durability after sudden power loss.

Validation: all 152 automated client tests passed before packaging. The grouped corrections were made in one client commit. Coverage includes damaged and missing files, strict device parsing, bounded FIFO reads, symlink targets, GPU settings preservation, CLI failures, unchanged Rusticl on invalid requests and desktop thermal saves. Tests use temporary homes and isolated children without registration, assignments or real search engines.

Coordination server: the deployed cancelled-range recycler remains unchanged. Larger cancelled linear ranges can still be split for smaller requests, preserving the remainder and checking selected intervals against protected work before reuse.

Privacy boundary: GPU readings remain in authenticated client heartbeats and the private Admin Nodes dashboard. The public website and network-status API expose no telemetry or physical GPU identifiers.

Package SHA-256: 8310e159e5dac5c72faa1ad3e1ec3b3a0ff63b16f1cef134b971951abf76e2d7

OpenPuzzle 1.0.44

One identity across concurrent first starts: cooperating clients use the new process lock when reading or creating their local identity. They agree on one complete UUID instead of generating different identifiers when they start together. Older clients do not participate in this lock.

Complete identity validation: truncated or malformed UUIDs and unexpected extra records are rejected before registration. Valid existing UUIDs retain their original text, including uppercase or lowercase hexadecimal digits and files with no line ending, LF or CRLF.

Preserved existing identities: only a missing identity file permits generation of a new UUID. Empty, malformed, unreadable or non-regular identity files are preserved and reported rather than silently replaced. Identity and lock symlinks are rejected without changing their targets, and FIFOs do not block the reader. The runtime reports the identity path when loading or creation fails.

UUID formatting independent of locale: first creation uses the classic numeric locale, so regional digit grouping cannot insert extra separators into the identifier. The established UUID version 4 layout remains.

No partial identity after failed creation: the new identity is published only after a complete private write succeeds. A failed write leaves no incomplete identity, removes only its temporary file and permits a later first-creation attempt. Identity and lock files are owner-only.

Recovery limits: damaged identity files are not repaired automatically. Preserve the original file and recover its UUID from a known backup or a matching local assignment before retrying. This update does not reconstruct an identity already lost, provide a transaction across multiple files or guarantee directory-entry durability after sudden power loss.

Validation: all 150 automated client tests passed before packaging. Five reproduced defects were corrected in one client commit. Regression coverage includes simultaneous first starts, process locking, failed writes, malformed identities, legacy line endings, locale grouping, non-regular files and symlink targets. Tests use temporary directories and request no assignments or engine searches.

Coordination server: the deployed cancelled-range recycler remains unchanged. Larger cancelled linear ranges can still be split for smaller requests, preserving the remainder and checking selected intervals against protected work before reuse.

OpenPuzzle 1.0.43

Complete configuration validation: malformed or incomplete JSON no longer applies some fields while leaving others at their defaults. The complete document is validated before its values are read, and invalid numeric or boolean fields retain their established defaults. Existing flat and nested configuration layouts remain supported.

Numbers independent of locale: saved numeric values use a stable JSON representation, including when the system locale uses decimal commas or grouped integers. Values with trailing text, fractional device identifiers and non-finite values are rejected rather than partially accepted.

Preserved files on write failure: failed writes or replacement operations preserve the previous configuration, client-state and execution files. This does not provide a transaction across multiple files or reconstruct data already lost.

Preserved client text: saved client-state strings retain carriage returns as well as newlines and literal backslashes when loaded again.

Conservative process identity: negative, overflowing or malformed saved process-start timestamps remain unknown. They are not treated as valid identities belonging to another process, and the assignment remains available to the existing identity and recovery checks.

Recovery independent of JSON presentation: recovery reads the stored fields correctly regardless of spacing or field order. A stored false value is not changed by the word true in a later command string. Invalid JSON is rejected without executing a partially recovered command.

Validation: all 149 automated client tests passed before packaging. Six reproduced defects were corrected in one client commit, with regression coverage for numeric configuration, client-state encoding, failed persistence and execution recovery.

Coordination server update: larger cancelled linear ranges can now be split to fit smaller work requests. The selected prefix and remaining suffix preserve the original interval without gaps or overlap. The remaining suffix is eligible for later work, and selected prefixes are checked against protected assignments and coverage before reuse. Splitting and assignment share one database transaction, so a failed operation preserves the original range. Independent Kangaroo sessions retain their existing routing.

OpenPuzzle 1.0.42

Confirmation before installing Kangaroo: installation requires the action to remain available after confirmation. If a runtime starts, its identity becomes uncertain or a status check fails while the dialog is open, confirming does not start the installer.

Confirmation before stopping: confirming Stop does not send a command if the client has become idle while the dialog is open. Existing stop retries remain available for occupied runtimes under the established runtime and identity checks.

Preserved thermal history: a malformed record preserves the original file and existing readings. Appending remains blocked until you explicitly clear the history, preventing a later reading from replacing the damaged file. Older records with missing optional fields remain supported.

Reliable configuration paths: saved paths preserve quotes, backslashes, control characters and Unicode when loaded again. Existing flat and nested configuration layouts remain supported.

Reliable execution recovery: saved execution details preserve escaped text, and recovery restores the original command contents. These fixes do not reconstruct data already lost.

Validation: all 146 automated client tests passed before packaging. Five reproduced defects were corrected in one commit, with new regression coverage for confirmations, configuration paths and execution persistence, plus expanded thermal-history checks.

OpenPuzzle 1.0.41

Reliable desktop launch: a status check started before clicking Start cannot re-enable searches or commands after the launch. The interface waits for a new status check instead of treating an older idle reply as an observation of the new launch.

Preserved launch controls: older replies, errors, crashes, empty or invalid output and timeouts preserve the pending launch and last confirmed status. Command buttons, execution choices and thermal configuration remain disabled. Language changes preserve these controls across all four interface languages. A timeout affects only the status check.

Refresh and recovery: after an older check finishes, the interface requests a new check automatically. A fresh valid response restores actions according to the observed runtime and identity state. A fresh query failure reports Status unavailable and blocks new searches until a valid response arrives. Foreground command protection from 1.0.40 remains in place.

Validation: all 143 automated client tests passed before packaging, including the new desktop launch-status ordering regression suite.

OpenPuzzle 1.0.40

Reliable desktop controls: new searches and command buttons remain disabled until the current command finishes or fails to start. Background status replies no longer enable overlapping commands. Execution choices and thermal configuration also remain disabled during the operation.

Live status: background status updates continue while a desktop command is running. Status errors and timeouts do not release that command, and a query timeout affects only the status check. Language changes preserve disabled controls across all four interface languages.

Completion and recovery: normal completion, nonzero exit, a process crash or failure to start release only the matching command. Available actions then follow the confirmed runtime and identity state: an active runtime keeps Start disabled and restores its stop controls; a confirmed idle client permits new work. Startup waiting and unavailable status warnings from version 1.0.39 remain in place.

Validation: all 142 automated client tests passed before packaging, including the new desktop command-reservation regression suite.

OpenPuzzle 1.0.39

Clear desktop status: the interface shows Status unavailable after a failed, crashed or timed-out status query, instead of treating the failure as an idle client. Empty, malformed or incomplete responses and partial output from a failed query do not confirm that work has stopped. New-work controls wait for the first valid status response.

Preserved confirmed data: last confirmed slot cards, assignment details, solution state and thermal readings/history remain intact. New searches, benchmarks, self-tests, Kangaroo installation and thermal configuration stay disabled while status is unavailable. The query timeout terminates only the interface's status subprocess.

Recovery and diagnostics: refresh and diagnostics remain available when the CLI can start. Stop controls retain their availability for previously confirmed occupied slots, with Kangaroo's safe-stop restriction. A later valid response restores the normal controls or the existing process-identity warning. The failure reason remains visible across all four interface languages.

Validation: all 141 automated client tests passed before packaging, including the new desktop status-failure regression suite.

OpenPuzzle 1.0.38

Clear desktop status: the desktop interface shows Identity unavailable when an execution cannot be confirmed, instead of displaying stopped or inferring running solely from a runtime PID. Affected slot cards stay visible, including primary, legacy and dynamic GPU workers and mixed active and uncertain execution states.

Protected desktop controls: new searches, benchmarks, self-tests, Kangaroo installation and thermal configuration stay disabled while any execution identity is unavailable. Refresh, diagnostics, update checks and auditing remain available. Stop controls can retry the CLI's identity-checked requests while retaining Kangaroo's safe-stop restriction.

Consistent recovery: a later confirmed running or idle status restores the established controls. Identity warnings and disabled controls remain consistent across all four interface languages.

OpenPuzzle 1.0.37

Clear process status: status and diagnostics report unavailable process identity instead of assuming that a runtime is idle or an engine has stopped. Uncertain identities are counted separately from confirmed stale controls, preserving local runtime and assignment state.

Reliable safe stop and updates: preserves existing safe-stop requests on repeated commands and partial failures. Primary and dynamic GPU workers are checked together, and installation is blocked while a runtime or engine identity is uncertain. Confirmed inactive identities retain their established behavior, including process exit, reboot and PID reuse.

OpenPuzzle 1.0.36

Protected runtime slots: blocks a new continuous runtime when an existing slot's process identity is temporarily unreadable or the current boot ID is unavailable. The existing runtime keeps its control marker, assignment state and safe-stop request instead of losing its slot to a second runtime.

Reliable startup decisions: checks again if another runtime appears during acquisition. A live or temporarily unreadable identity blocks takeover, while confirmed stale markers still permit startup after a process exits, the machine reboots or a PID is reused. Other execution slots retain their own control and requests. Failed-stop preservation from version 1.0.35 remains in place.

OpenPuzzle 1.0.35

Reliable stop control: preserves runtime control and local assignment state when a stop request fails. An active runtime keeps its local control marker, preventing a second runtime from acquiring the same execution slot while the first is still running.

Clear failure reporting: unreadable process identities preserve runtime control. The command reports the affected execution slot and returns an error when a stop cannot be delivered. Other GPU slots receive independent stop requests, and partial failures preserve the affected assignment instead of claiming that every runtime is shutting down.

OpenPuzzle 1.0.34

Reliable recovery: recovers interrupted assignments even while a new runtime is active, once the saved process is confirmed stopped. An absent process, a reused PID or a different boot identifies an old execution that needs recovery. This fixes a loop that could prevent the client from requesting new work.

Conservative synchronization: unreadable process identities and temporary network failures preserve local state for retry. Interrupted ranges are not reported as completed. Accepted cancellation or explicit assignment rejection clears only the affected execution state, keeping recovery independent for each execution slot and preserving protected local results.

OpenPuzzle 1.0.33

During multi-GPU execution, a surviving worker takes over private GPU telemetry collection if its previous owner exits, continuing its own assignment without restarting a GPU worker.

The Qt interface preserves an unreadable local history file and shows an actionable error instead of replacing the file with an empty history.

OpenPuzzle 1.0.32

Added a persistent local thermal history of GPU Warning, Critical, Invalid and recovery transitions. The newest 200 events survive interface restarts and can be cleared after confirmation. Protection remains disabled by default.

OpenPuzzle 1.0.31

Added persistent local desktop alerts for active GPUs in Warning, Critical or Invalid thermal state, identifying the runtime slot and temperature. State transitions and recovery are announced once; critical alerts report whether orderly stop was requested or monitoring is diagnostic-only.

OpenPuzzle 1.0.30

Added desktop controls for the optional thermal policy and a localized thermal-state badge to every active GPU card while retaining the existing command-line configuration and disabled-by-default protection.

OpenPuzzle 1.0.29

Added optional per-GPU temperature and power reporting to authenticated heartbeats, the private Admin Nodes dashboard and matching Qt runtime cards while keeping the public website and network-status API telemetry-free.

OpenPuzzle 1.0.28

Added read-only NVIDIA and AMD telemetry, an optional diagnostic thermal policy and explicit orderly protection at a configured critical threshold. Monitoring remains scoped to the GPUs selected for each execution.

OpenPuzzle 1.0.27

Added unlimited multi-CUDA execution with an independent dynamic cuda-N supervisor, assignment, workspace and recovery lifecycle for every selected device. Cached discovery and CUDA device isolation made larger rigs more efficient while retaining the existing runtime controls.

OpenPuzzle 1.0.26

Added the reproducible, self-validating Linux release pipeline covering the Debian, portable, binary and exact-source packages, provenance metadata, engine identities and updater manifests.

OpenPuzzle 1.0.25

Added the optional multilingual Qt 6 desktop controller with light and black-and-gold themes, official branding, multi-slot status and direct controls for the audited command-line client.

Made runtime inspection strictly local and read-only, added exclusive Kangaroo execution for supported puzzles and retained independent BitCrack CUDA/OpenCL and KeyHunt CPU operation.

OpenPuzzle 1.0.23

Added complete backend-aware GPU inventory reporting, including AMD devices exposed through Rusticl, and accurate capability reporting for BitCrack, KeyHunt and Kangaroo.

OpenPuzzle 1.0.22

Added solution-review control that stops work while a puzzle is awaiting verification or after it is confirmed as solved, without transmitting private-key material or local solution exports.

Improved monitoring of transient state failures and added duplicate-engine protection based on Linux boot identity and process start time.

OpenPuzzle 1.0.21

Added protected Pollard Kangaroo checkpoint recovery with explicit non-zero walk seeds, a verified reproducible PSCKangaroo build path and independent CUDA/OpenCL service-slot configuration while preserving explicit BitCrack and KeyHunt routing.

OpenPuzzle 1.0.20

Added metadata-driven routing for linear and Kangaroo-compatible puzzles, external PSCKangaroo integration contracts and resilient metadata-only solution alerts while keeping private-key material entirely local.

OpenPuzzle 1.0.19

Preserved Rusticl OpenCL driver selection across client commands and added verified release manifests for safer package discovery and installation.

OpenPuzzle 1.0.18

Extended Linux process identity to include boot ID, PID and process start time, preventing both cross-reboot and same-boot PID reuse from being trusted as an active OpenPuzzle execution.

OpenPuzzle 1.0.17

Bound runtime state, background supervisors, heartbeat, stop handling and update recovery to the current Linux boot identity.

OpenPuzzle 1.0.16

Added resilient recovery for executions interrupted by reboot, power loss or abrupt shutdown, preserved local state across temporary API failures and kept CUDA/OpenCL recovery independent. It also established the verified safe-update path used by later releases.

OpenPuzzle 1.0.10

Added a wider bounded GPU benchmark matrix, actionable error codes, protected visible solution notices and the verified openpuzzle update workflow.

OpenPuzzle 1.0.9

Added local backend self-tests for CUDA, OpenCL and CPU execution, plus packaging and runtime validation improvements.

Earlier 1.0 releases

The first public 1.0 series established anonymous continuous assignments, progress and completion reporting, CUDA/OpenCL/CPU packaging, safe stop, recovery and protected local solution handling. Historical tags and source remain available in the public repository.