AutoForge/docs/build-and-test.md

953 lines
60 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

# Build and test
## Prerequisites
- Windows x64, .NET SDK 10.0.402 with its approved .NET 10 patch roll-forward.
- Windows SDK target 10.0.26100.0 and MAUI Windows workload manifest 10.0.110.
- Microsoft Edge WebView2 Runtime to launch the native Blazor Hybrid Client.
The previously verified environment reported runtime 10.0.13, MAUI workload
version 10.0.403 as listed by the installed SDK, and WebView2 Runtime
155.0.4283.62. The Client targets `net10.0-windows10.0.26100.0` and `win-x64`.
All other projects target portable `net10.0`; tests do not require MAUI or
WebView2.
## Restore, build, test and launch
From the repository root in PowerShell, run the full build and both suites
sequentially:
```powershell
dotnet restore .\src\AutoForge.slnx
dotnet build .\src\AutoForge.slnx --configuration Release --no-restore
dotnet test .\src\AutoForge.slnx --configuration Release --no-build
dotnet run --project .\src\AutoForge.Client\AutoForge.Client.csproj --configuration Release --framework net10.0-windows10.0.26100.0
```
The build includes the actual native Client and all nine solution projects.
For an explicit Windows runtime pack restore or direct native build:
```powershell
dotnet restore .\src\AutoForge.Client\AutoForge.Client.csproj --runtime win-x64
dotnet build .\src\AutoForge.Client\AutoForge.Client.csproj --configuration Release --framework net10.0-windows10.0.26100.0 --no-restore
```
Do not pass Windows TFM overrides to the portable test projects. The Client
contains the real UI RCL in a native MAUI BlazorWebView. No automotive driver,
actual vehicle, database, backend, model or cloud account is required.
## Focused tests
```powershell
dotnet test .\src\tests\AutoForge.UnitTests\AutoForge.UnitTests.csproj --configuration Release
dotnet test .\src\tests\AutoForge.IntegrationTests\AutoForge.IntegrationTests.csproj --configuration Release
dotnet test .\src\tests\AutoForge.IntegrationTests\AutoForge.IntegrationTests.csproj --configuration Release --no-restore --filter "FullyQualifiedName~TelemetryReplayTests|FullyQualifiedName~LiveDashboardAcquisitionTests|FullyQualifiedName~DisplayUnitControlTests"
dotnet test .\src\tests\AutoForge.IntegrationTests\AutoForge.IntegrationTests.csproj --configuration Release --no-restore --filter "FullyQualifiedName~CalibrationAuthoringTests"
dotnet test .\src\tests\AutoForge.IntegrationTests\AutoForge.IntegrationTests.csproj --configuration Release --no-restore --filter "FullyQualifiedName~CalibrationRehearsalTests|FullyQualifiedName~RenderedRehearsal|FullyQualifiedName~RehearsalIsolation|FullyQualifiedName~CalibrationAuthoringTests"
dotnet test .\src\tests\AutoForge.IntegrationTests\AutoForge.IntegrationTests.csproj --configuration Release --no-restore --filter "FullyQualifiedName~CalibrationRehearsalReopenTests|FullyQualifiedName~CalibrationRehearsalTests|FullyQualifiedName~RenderedRehearsal|FullyQualifiedName~RehearsalIsolation"
dotnet test .\src\tests\AutoForge.IntegrationTests\AutoForge.IntegrationTests.csproj --configuration Release --no-restore --filter "FullyQualifiedName~RunbookPreviewTests|FullyQualifiedName~RunbookUi_|FullyQualifiedName~RunbookIsolation_"
dotnet test .\src\tests\AutoForge.IntegrationTests\AutoForge.IntegrationTests.csproj --configuration Release --no-restore --filter "FullyQualifiedName~LocalRunbookGenerationTests|FullyQualifiedName~LocalInferenceBoundsTests|FullyQualifiedName~LocalGenerationReviewTests|FullyQualifiedName~LocalGenerationUi_|FullyQualifiedName~LocalGenerationIsolation_|FullyQualifiedName~RunbookPreviewTests"
dotnet test .\src\tests\AutoForge.IntegrationTests\AutoForge.IntegrationTests.csproj --configuration Release --no-restore --filter "FullyQualifiedName~Comparison_|FullyQualifiedName~PreviewHandoff_|FullyQualifiedName~TorqueDraft_"
```
## Step-12A real local-model/native qualification
The corrected owner-directed handoff authorized scoped local-only setup without
the obsolete approval form. The current Ollama 0.40.13 startup now reports
`Ollama cloud disabled: true` (2026-10-31 14:39:16 -04:00), with the exact installed
model/digest/store/version and executable hashes preserved.
Native original exact-goal output passed the unchanged validator with only
+600 rpm/-10 N*m and unchanged unrelated cells/axes/source. Actual pending
conflicts/Cancel/unmount/fresh owner, manual/stale/units and source preservation
passed; counted controlled coexistence remains separately identified.
Native testing exposed a Hybrid event/replaceable synchronous-transport prefix
blocking pending rendering. Two pre-fix corrective cases reproduced it. UI
callbacks now release IPC events and observe/render outcomes; admitted owner
work runs off the UI dispatcher, with source/model/preview/cancellation guards
unchanged. Focused generation/validator selection: 151 passes (49 local-generation
cases plus 102 parser cases), zero failed/skipped.
Final sequential Release build: all nine projects, exit 0, 0 warnings/errors.
Full tests: exit 0, 179 unit + 417 integration = 596 passes, zero failed/skipped;
all 594 prior regressions retained plus two corrective cases.
See [Step-12A execution evidence](../prompts/Step-12A-Execution-Report.md).
Step-12/12A is accepted.
## Step-13 verified comparison verification
The focused comparison/pinned-draft selector passes 50 cases, including actual
owning proof, whole-config equality, exact nonuniform/tiny canonical handoff,
reconciled Apply, malformed/unavailable fresh target reads, real wire signal
faults, measured 3900/3901 and fewer-than-ten-warm boundaries, held actual
transport/final-read/save Cancel/Close/source supersession, partial/uncertain
publication retry and real file/readback/replay corruption.
The original full-scenario regressions still independently check all 3000 cruise
model steps; production records 600 sampled cruise signal sets, not continuous
extrema. Native demonstration includes the existing real local model's original
goal/output, separate draft/Save/backup/approval/Apply/full readback, equivalent
3361-step real pair, explicit two-capture readback and existing RecordedReplay
raw trace. Controlled faults/barriers are not relabeled as native physics/model
proof. See [comparison usage](calibration-comparison.md) and
[Step-13 execution evidence](../prompts/Step-13-Report.md).
Step-13 awaits manual review; no Step-14 or runtime installation/setup is implied.
Final sequential Release build: nine projects, exit 0, zero warnings/errors.
Full tests: 179 unit + 466 integration = 645 passes, zero failed/skipped,
retaining all 596 original cases and adding 49 comparison/handoff cases.
The unchanged warm-baseline regression now independently asserts exactly
3361 samples/336.1 s and every one of the 3000 cruise model steps within
45 MPH +/-1 MPH, rather than using the production sampled range as a proxy.
One intermediate full run failed in the unmodified
`ScenarioOwnerCannotBeReplacedAndResetOrCloseSuppressesOldRunPublication`
cancellation regression; the isolated seven-case owner selector and subsequent
full runs passed. This intermittent result is recorded, not concealed or fixed
outside the assignment. Rapid native diagnostic/simulation navigation also
exposed the unchanged recording panel's pending-refresh Busy initialization
error. Task-owned Client recovery and sequential exact-heading/idle-archive-owner
checks passed for all six existing routes. No recording-navigation change was made.
## Historical Step-12 approval-blocked verification
Step-11 is accepted. Step-12 implementation/control tests pass, but required
real-model/native exact-goal success is **NOT EXECUTED/BLOCKED**, not waived.
The original runtime inspection reported Ollama 0.40.13 and an empty model
catalogue. Completion inspection now verifies the operator-installed
`huihui_ai/qwen3-abliterated:8b-v2` and exact approved digest/local details/license.
The current server startup log still reports `Ollama cloud disabled: false`.
Specific configuration/restart approval and quiescence confirmation were
unavailable, so the completion worker made no setup change, restart, download or
chat request. No goal/source context was sent to that shared runtime.
The focused selection passed 149 cases (47 new local-generation cases plus
102 manual-parser cases). The complete nine-project Release build passed with
0 warnings/errors, followed by 179 unit + 415 integration = 594 passes,
zero failed/skipped, preserving all 547 baseline cases.
Controlled HttpMessageHandler/streams test actual request/response contracts,
local eligibility, original output, exact +600/-10 canonical diff, HTTP/depth/
critical-field errors, 131072/131073 request bytes, 1048576/1048577 actual
lengthless response bytes and controlled 180-second cancellation. Barriers prove
source/model/manual-preview supersession, no stale overwrite and cancel/unmount
drainage. Rendered controls and real business-source/recording/replay/approved-
rehearsal isolation verify zero unassigned mutations. These are not real LLM
intent/native-generation evidence.
Historical native checks verified no automatic request/selection/acknowledgement on entry,
explicit real empty-model catalogue, visible ModelRequired rejection before
context/chat and continued manual NoChanges preview/source/old-route behavior.
Cloud-disabled acknowledgement was deliberately left false. Runtime configuration
was not altered. Read-only live prerequisite inspection and controlled HTTP
tests are distinguished from model-generated output in the
[original Step-12 report](../prompts/Step-12-Report.md) and
[local generation guide](local-runbook-generation.md).
Completion native checks on the actual Client verified explicit exact-model
selection (same name/digest/5,027,793,060-byte local GGUF), current warm-baseline
source/head and matched knowledge inspection without chat. Acknowledgement
remained false and Generate visibly rejected `LocalOnlyAcknowledgementRequired`
before dispatch. Manual NoChanges and invalid-input retention, unchanged
canonical source/head/history, idle-owner close and fresh re-entry passed.
These are not real generation, pending Cancel/unmount or exact-intent evidence.
The required sequential Release build again returned 0 warnings/errors; tests
again passed 179 unit + 415 integration = 594, zero failed/skipped. No application
or test code changed. See the separate
[Step-12 completion report](../prompts/Step-12-Completion-Report.md) for the precise
approval blocker and prepared operator action.
Step-12A freshly rechecked exact installed-model metadata/resources and the
same current cloud-enabled startup. Explicit setup/restart approval and idle
confirmation again returned unavailable, not consent. Per its setup-blocker
rule, the worker did not launch native qualification or rerun build/tests only
to restate the permission blocker. Thus the above 594-case/clean-build results
remain historical Completion evidence; Step-12A build/test/native qualification
is NOT EXECUTED/BLOCKED. See the [Step-12A report](../prompts/Step-12A-Report.md).
## Step-11 deterministic runbook preview verification
Corrected Step-10/10A and Step-11 are accepted. Historical Step-11 evidence follows.
The focused runbook command passed **105 cases**, including exact source/head/
candidate/diff vectors, every schema object's missing/duplicate/unknown fields,
wrong kinds/integers/allowlist/identity/unit/axis/full-shape vectors, defensive
copies, stale result retention, safe rendered rationale and no auto validation.
Input at exactly 131072 UTF-8 bytes succeeds; one more fails without replacing
the last preview. Rationale 2000/2001 and step-ID 128/129 UTF-16 boundaries,
supplementary Unicode, invalid encoding/escapes and parser depth 32/33 are
asserted separately. Tiny canonical differences are not rounded away.
The required sequential nine-project Release build passed with **0 warnings,
0 errors**; the complete suite passed **179 unit + 368 integration = 547 cases**,
zero failed/skipped, preserving all 442 baseline cases. Real live diagnostics/
dynamics, saved recording, validated replay and a pending approved rehearsal
remain unchanged across template/valid/invalid preview invocation. Counted
transport/model/archive-Save/rehearsal open/read/write deltas are zero; authoring
change notifications and history mutations are zero, with unchanged approval,
source export, trace/state, archive bytes and replay snapshot.
Native smoke on the rebuilt Client verified empty-source rejection, explicit
source/template with no auto preview, NoChanges, exact 650->1250/+600 rpm and
360->350/-10 N*m candidate/diff, safe rationale text and display-only units.
Malformed, physical-target, wrong-head and invalid-value inputs visibly rejected
while retaining the previous preview ID and exact diff. An explicit authoring
head change marked that captured result Stale; old input rejected and a new
template still did not validate until the explicit action. Current Proposed
head revalidation produced NoChanges without another authoring version.
Source export stayed identical and original preferences/old routes worked.
This normal UI evidence is not headless exact-limit/coexistence proof.
See the [Step-11 worker report](../prompts/Step-11-Report.md).
## Step-10A pending-reopen close ownership verification
Corrected Step-10/10A is accepted. Original corrective results below and the
worker reports are preserved.
Before the fix, the two `PendingReopenCloseOrDispose_DrainsCurrentOwnerWithoutPriorCompletionReplay`
cases executed against the actual Simulation factory and failed with:
prior close reused=True, Close/Dispose completed while factory held=True,
released Open error=none, state=Ready and new target dispose count=0.
Barriers were released and actual targets drained in finally, so the test
runner exited with two behavioral failures rather than stranded workers.
After the correction, all six new cases passed. They assert a new current-close
completion, cancellation/drain while factory return and actual target cleanup
are held, no Ready publication, retained inactive evidence, exactly one real
target disposal and unusable disposed sessions. Prior failed close, current
factory/cleanup failure, rejected reopen and later valid sequential reopen
are exercised. The combined corrective/original rehearsal selector above
passed **37 cases**, zero failed/skipped.
The required sequential nine-project Release build passed with **0 warnings,
0 errors**, followed by **179 unit + 263 integration = 442 passed**,
zero failed/skipped. All 436 baseline cases remain unchanged.
The rebuilt native Client rechecked explicit source/proposal/target selection,
two-read backup, confirmed Apply/full readback, separately confirmed original
Restore, repeated Close, new-instance/zero-revision reopen and calibration
route close/re-entry. A combined proposal stored 1250 rpm and 350 N*m at
2000 rpm in the isolated image at revision 1; original Restore recovered
650/360 and every other cell/axis at revision 2. Source export/fingerprint
remained identical; a separate ordinary dynamics read at step 7 returned
631.80416 rpm from the unchanged model. Units preserved approval without
dispatch, and original preferences/old replay route remained intact.
The pending-factory race is controlled test proof, not native timing evidence.
Exact commands, observations, cleanup and limitations are in the
[Step-10A corrective report](../prompts/Step-10A-Report.md).
## Step-10 simulated programming rehearsal verification
Step-09 and corrected Step-10/10A are accepted. The historical focused
Step-10 command passed **86 cases**: 31 new rehearsal cases plus 55 accepted authoring
cases. They use actual Simulation target storage/CAS through typed Diagnostics
ports, real authoring/configuration/profile services and controlled forwarding
faults. Independent backup, complete scalar/table readback, immutable approval,
stale head/target CAS, cancellation before/after actual commit, exact repeated
operations, negative/echoed/lost acknowledgements, divergent/missing reads,
read-only reconciliation, separately approved Restore, held close/failure
drainage and exact 32/33 audit capacity are verified.
The required sequential Release build completed all nine projects/native
Client with zero warnings/errors. The full suite passed **179 unit +
257 integration = 436 cases**, zero failed/skipped, preserving all 405
baseline cases. Real live dynamics/diagnostics, saved recording and validated
replay coexist while rehearsal adds zero calls to their transport/model/archive
Save boundaries and leaves configuration export, live trace/status, archive
bytes, authoring history and replay snapshot unchanged.
Native smoke used the actual rebuilt Windows Client and task-owned WebView2
profile/local DevTools port. Within `/calibration`, an explicitly created warm
configuration/proposal opened the distinct Simulation rehearsal target. Two
independent reads produced its verified supported-image backup at revision 0.
Explicit scalar Apply yielded 1250 rpm with all torque cells unchanged at
revision 1. A new proposal changed only 2000-rpm torque 360->350 N*m; its new
approval/full readback yielded revision 2. Separately confirmed original Restore
yielded 650 rpm and every original torque cell/axis at revision 3. Close left
inactive evidence; reopen created a different instance with no old backup/
approval/audit. Restore without a fresh backup rejected.
The native source export/fingerprint remained identical, with source warm idle
650 rpm and 2000-rpm torque 360 N*m. A separate actual dynamics read at model
step 7 returned 631.80416 rpm, not the rehearsal's 1250 image. Display units
preserved an unconsumed approval without dispatch; exact repeat dispatched no
second mutation. Original Metric/Road preferences and old replay/home routes
worked. Initial fast automation hit stale-head/DOM-readiness rejections; exact
selected/current-head feedback and settled real controls were then used for
the complete successful workflow. Those are not controlled race/fault proofs.
See the [Step-10 worker report](../prompts/Step-10-Report.md) for exact normal/
controlled evidence, commands, cleanup and preservation.
## Step-09 synthetic authoring verification
Steps 08 and 09 are accepted. The historical Step-09 worker evidence follows.
The focused authoring command above passed **55 cases**, including real
configuration/profile source resolution, exact 650->1250 rpm and
360->350 N*m diffs/parents, changed-back/no-op/tiny changes, defensive copies,
template/source rejection, controlled competing-head commits and the exact
32/33 history boundary. Real rendered events verify parse/validation failures,
history/diff selection, two stale editors, explicit replacement and pinned
torque units with exact canonical preservation.
The required sequential Release build completed all nine projects/native
Client with zero warnings/errors, then the complete suite passed **179 unit +
226 integration = 405 cases**, zero failed/skipped. All 350 accepted baseline
cases remain present. A real live dynamics source, its saved recording and
validated replay coexist during authoring with zero additional transport,
model-control and archive Save calls, unchanged source export/fingerprint/
trace/state/raw bytes and unchanged replay snapshot.
Native smoke used the actual Release Windows Client and a task-owned WebView2
profile/local DevTools endpoint. The existing simulation workflow explicitly
created the warm source. Native authoring created its baseline, saved the
1250-rpm scalar proposal and 350-N*m torque cell at 2000 rpm, then inspected
exact baseline and preceding-version diffs. Invalid scalar/cell/source commands
and unconfirmed replacement preserved the head/history. A Metric draft's
unsaved 345 N*m remained 345 after an Imperial toggle; cancelling captured a
new Imperial draft, and an untouched save across another toggle retained
exact source 350 N*m and the same revision. Confirmed replacement explicitly
discarded the old draft/history and returned to baseline.
The original dynamics configuration export/fingerprint stayed identical with
warm idle 650 rpm and 2000-rpm torque 360 N*m. An explicit subsequent real
diagnostic read at model step 7 / 0.7 seconds returned 631.80416 rpm from the
unchanged warm model, not the proposal. Old simulation/profile/replay routes
and authoring re-entry worked. Restart confirmed no authoring workspace or
configuration persisted; the final binary then repeated the baseline and
two-proposal diff workflow. Native observations are not controlled concurrent/
stale-head/capacity proof. Exact commands, final results, cleanup and limitations
are in the [Step-09 worker report](../prompts/Step-09-Report.md).
## Step-08 recorded replay verification
Corrected Step-07/07A/07B and Step-08 are accepted. The recorded Step-08 focused
command passed 49 cases. It covers
37 new replay cases plus accepted live-dashboard/unit-selector regressions.
The required sequential nine-project Release build passed with zero warnings
and errors, followed by 179 unit and 171 integration tests: **350 passed,
zero failed/skipped**, preserving all 313 baseline cases.
The replay tests construct real recorder/codec/file-store inputs, then dispose
their original sources and load without profile/configuration catalogues.
Independent four-observation assertions cover exact prefixes, last-Good versus
Uncertain, ties, before-first, zero/empty/non-signal archives and model reset/
repeated time. Controlled monotonic waits/barriers prove 0.25x/1x/1000x pacing,
1.5-to-1.7-second mid-gap pause at 2x, remaining 0.15-second resume, held cleanup,
fault recovery (including faults during pending Pause/Stop cleanup), stale
loads/wakes and unmount. Actual rendered route/editor
events prove committed selector feedback, persistent actual fault reasons despite
older command errors, partial acknowledgement, pinned units and atomic edits.
A real simultaneously open dynamics source retains
its trace, model state and unsaved recording with zero additional transport,
model-control or archive Save calls during replay. Shared gauge hide/units/
unmount also leaves an active replay owner untouched.
Native smoke uses the actual Windows Client through a task-local WebView2
DevTools endpoint, not a second web host. Saved static-v1 and dynamic-v2 captures
were recorded through native controls, stopped, source-closed and saved.
After app restart, the profile catalogue was empty; both files replayed without
reimporting profiles or recreating a model/configuration. Native checks include
explicit load/no autoplay, numeric/bar gauges, Play/Pause/Stop/First/Restart,
time/ordinal/backward navigation, rejected speed feedback, hide/show, end,
route close/re-entry, raw/decoded inspection and restored Metric/Road
preferences. A recorded 88.02090573589594 degC source showed 88.02 degC or
190.44 degF; draft 32..212 degF saved without changing canonical 0..100 degC
bounds. Time seek 40,000,000 ticks retained exactly two of four dynamic records.
The native partial warning/acknowledgement/end check used a separately named,
strict-reader-validated derived static fixture; it is not evidence that native
recording exhausted a limit. Actual limit/error capture and controlled timing/
fault/interleaving evidence remain in headless tests. Static/dynamic archive
SHA-256 hashes were unchanged after playback. Exact task-owned capture IDs,
storage path, cleanup and native preparation limitations are recorded in the
[Step-08 worker report](../prompts/Step-08-Report.md).
MAUI storage was verified at its real app-local directory, not inferred from
`LOCALAPPDATA`. Only individually named task-owned capture files and task-owned
WebView profile/helper were cleaned up; unrelated app data/preferences remain.
The unit suite covers existing Core/Application behavior, rendered generic
components, profile invariants and strict codec behavior, plus protocol v1
round-trip/snapshot contracts and bounded trace behavior. The integration suite
loads source-controlled profile fixtures through real DI, opens actual
simulation sessions, sends/decodes requests through the transport, verifies
identity/DTC/static signals, virtual clear/readback/reset and independent
session state, exercises deterministic malformed/no-response/timeout/cancelled
faults, renders the diagnostic route against the actual provider, and checks
the nine-project reference graph. Fixtures are never treated as real vehicle
definitions. The internal state oracle is exposed only to the integration-test
assembly.
The accepted Step-02A baseline was 103 unit plus 37 integration tests (140
total). The Step-03 worker report records the full, expanded counts, warnings,
commands and native observations; do not infer acceptance from these earlier
runs. Historical Step-02 had 59 unit and 34 integration cases; it predates the
Step-02A corrections. See the [Step-02 report](../prompts/Step-02-Report.md) and
[Step-02A report](../prompts/Step-02A-Report.md) for historical evidence.
Step-03 worker verification ran the nine-project Release build with zero
warnings/errors, then passed 108 unit and 53 integration cases (161 total,
zero failed/skipped). Its native smoke used the actual Windows Client and
verified explicit synthetic selection, module probes/identity, DTC and static
signals, raw request/reply trace, virtual clear/readback/reset, close, target
metadata, profile, home and not-found routes. Exact observations and limitations
are recorded in the [Step-03 worker report](../prompts/Step-03-Report.md);
the run did not cover the subsequently identified integrity findings and Step-03
was not accepted.
Step-03A added actual failing-before/passing-after regressions for missing and
duplicate request fields, semantic response validation/trace consistency, and
clear acknowledgement versus verified readback. The final sequential Release
verification passed 111 unit tests and 64 integration tests (175 total, zero
failed/skipped), preserving the original 161 cases plus 14 additional test
cases. Exact commands and outcome details are in the
[Step-03A corrective report](../prompts/Step-03A-Report.md). Manual
orchestrator acceptance remains pending.
Step-03B reproduced the probe-only identity-loss defect before the fix through
the real Simulation transport: the regression expected a positive probe but
observed a negative discovery outcome. After the correction, the focused
protocol/UI integration selection passed 35 cases. The complete nine-project
Release build passed with zero warnings and errors, and the complete suite
passed 111 unit plus 72 integration tests (183 total, zero failed/skipped),
preserving all 175 Step-03A cases. Exact commands and AC mapping are recorded in
the [Step-03B corrective report](../prompts/Step-03B-Report.md).
The native Windows Client was launched and the operational baseline was
exercised in its own WebView: explicit example import/profile/revision/provider/
Simulation selection, session open, discovery with expected/probe/identity
counts, configured transmission no-response, engine DTC/static-signal reads,
verified clear with fresh empty readback, raw/decoded trace, reset with initial
DTC restoration, close, and `/execution-targets`, `/vehicle-profiles`, `/`,
and not-found routes. Exact observations and limitations are in the
[Step-03B corrective report](../prompts/Step-03B-Report.md). Native interaction
used a session-local WebView debug port on the worker-launched Client; partial
identity remains covered by real-DI rendered tests because the shipped native
configuration has no fault selector for that scenario.
## Step-04 fixed-step dynamics verification
The Step-04 Release regression ran sequentially against the full solution:
```powershell
dotnet build .\src\AutoForge.slnx --configuration Release --no-restore
dotnet test .\src\AutoForge.slnx --configuration Release --no-build
```
The native Windows-targeted build covered all nine solution projects with zero
warnings/errors. The complete test command passed 115 unit and 75 integration
tests (190 total, zero failed/skipped), preserving the Step-03B 183-test
baseline. The unit numerical gate executes 3,361 deterministic 0.1-second
samples and compares full sampled state across equal runs. The baseline results
were: 10.0 s warm idle at 650.000 rpm; acceleration to 20.1473 m/s in 10.3 s
with four scheduled shifts; all 3,000 cruise samples inside the configured
45-MPH +/- 1-MPH band and fully coupled; braking to 0.0000 m/s in 5.8 s; 10.0 s
return idle; and 0.518387 kg modeled total fuel. These are synthetic model
outputs, not measured vehicle performance or manufacturer predictions.
The actual Windows Client was launched with WebView2 local DevTools enabled for
the native smoke. `/simulation` required explicit selection of profile
`autoforge.synthetic.test-truck` revision 1, provider
`autoforge.simulation.dynamics`, and a configuration with fingerprint
`sha256:v1:e8991ae139b54959ec53a4fe64b08dc9bfc03feb533496cf741f2c9f3a2342da`.
Export/import preserved that fingerprint. A stopped open session returned 17
protocol-v2 dynamic telemetry values at step 4 / 0.4 model seconds after explicit
start/pause/step controls. The accelerated full scenario completed at step
3,361 / 336.1 model seconds with 17 telemetry values and four shifts. Invalid
`{}` import visibly reported the missing required `formatVersion`; exact-config
reset opened a new session at step 0 with the same fingerprint, and close
released it.
The native `/diagnostics` regression separately selected the original
`autoforge.simulation.synthetic` provider, discovered two positive probes out
of three expected endpoints (the transmission remained configured no-response),
and read the original engine DTC plus static 850-rpm and 12.6-V signals through
protocol v1 with trace records. The session was closed. Native route smoke
covered `/execution-targets` (both providers remain metadata), `/vehicle-profiles`,
`/`, and an unmatched not-found route. Numerical fidelity and reproducibility
are established by headless tests, not by visual telemetry.
## Native synthetic diagnostics smoke
Launch the actual Windows Client and visit `/diagnostics`:
1. Confirm the page starts without selecting/importing a profile or opening a
session. Click **Load synthetic truck example**; verify the profile appears
as an available revision but remains unselected.
2. Explicitly select `autoforge.synthetic.test-truck`, revision `1`, provider
`autoforge.simulation.synthetic`, and kind `Simulation`. Click **Open
simulation session** and confirm the visible session/profile/provider/
configuration identity. The session opens in process only.
3. Click **Probe virtual modules**. Engine/body respond; the configured
transmission endpoint visibly reports no response and is not reported as an
actual responder. Inspect the identity and declared implemented services.
4. Read the engine's DTCs and static signals. Values are synthetic static
snapshots with source, unit, quality and logical tick. Expand the bounded
trace to inspect actual request/response JSON and correlation values.
5. Clear the supported engine virtual DTCs, confirm the fresh readback is empty,
and confirm the UI explicitly states that no physical write occurred. Other
module/session state is isolated.
6. Exercise a visible deterministic failure (for example the configured
transmission no-response), then reset to the initial synthetic state and
close the session. Closing releases that session's trace and virtual state.
7. Separately visit `/execution-targets`, `/vehicle-profiles`, `/`, and an
unmatched route. Confirm metadata selection remains separate from an
operational session and the not-found/navigation behavior remains.
The Step-03A report records actual native baseline smoke separately from
headless rendered fault tests. The native Client revalidated explicit
profile/provider selection, successful clear with fresh empty readback,
trace/raw frames, visible no-response, reset/initial-state restoration, close
and existing routes. A deliberately false clear acknowledgement was injected
only in the rendered integration composition; the production Client has no UI
fault selector, so this is not native fault-injection evidence.
## Profile-schema and safety boundaries
See the [schema-v1 guide](vehicle-profile-schema-v1.md) for field, identity,
unknown-value, capability, provenance and strict JSON rules. Important invariants:
- Every v1 field is required; unknown optional facts are explicit JSON `null`.
Unknown fields and duplicate/case-ambiguous properties are rejected.
- Catalog selection requires exact `(profileIdentifier, revision)`. Duplicate
imports do not replace entries, and failed parsing cannot partially register.
- `publishedAt`/`retrievedAt` accept null or complete
`YYYY-MM-DDTHH:mm:ss[.fraction](Z|+HH:mm|-HH:mm)` timestamps only. The exact
fractional and offset limits and no-default policy are in the schema guide.
- A null optional module identity in evidence is an absent constraint. Any
supplied identity must match a known module value; an unknown target or
mismatch is rejected without identity inference.
- Imported profiles and capability/provenance statements are immutable
metadata, not connection, compatibility, a working service or authorization.
- Data remains in memory for the process only. JSON import/export does not write
files or timestamps implicitly.
- Static synthetic diagnostics use the [protocol v1 guide](synthetic-diagnostics-v1.md)
and [simulation definitions](simulation.md); dynamic output uses the separate
documented protocol-v2 provider. Neither is a real vehicle adapter, telemetry
capture, replay provider, calibration/firmware editor or hardware write path.
## Step-04A corrective verification
Step-04A corrects typed import errors for unavailable profile/revision/module or
conflicting calibration references, enforces the supported coupled RPM/redline
envelope, and makes progress, terminal failure and asynchronous runner ownership
explicit. It adds seven unit and five integration cases over the 190-case
Step-04 baseline. The final sequential commands were:
`dotnet build .\src\AutoForge.slnx --configuration Release --no-restore`
`dotnet test .\src\AutoForge.slnx --configuration Release --no-build`
The nine-project Windows Release build succeeded with zero warnings and zero
errors. The full test command passed 122 unit and 80 integration tests (202
total; zero failed/skipped). The warm baseline remains reproducible at 3,361
samples / 336.1 virtual seconds; 10.0 s warm idle at 650 rpm, 10.3 s acceleration
to 20.1473 m/s with four shifts, all 3,000 cruise samples in the original
45-MPH +/- 1-MPH band while coupled, braking to 0.0000 m/s in 5.8 s, 10.0 s
return idle, and 0.518387 kg modeled fuel. The added independent constant-torque
force/mass equation and speed-step equation checks use a `1e-10` acceleration
tolerance; independent locked RPM checks use `1e-8` rpm.
Controlled-pacer unit tests exercised active pause/freeze, resume, stop, reset,
restart, cancellation and disposal without wall-timer polling. Equivalent full
scenario samples were identical at 0.25x and 1000x; only requested pacing delay
changed. Rendered integration tests covered typed/atomic profile, revision and
module import failures; a persistent 0.1-second deadline failure and negative
`DYNAMICS_RUN_FAULTED` response; UI run reentrancy; and stale-run suppression
across reset, close, reopen and delayed open.
Headless numerical/integration coverage also accepts a finite but extreme
torque/gear vector that overflows modeled power during launch. The engine
rolls back to its last finite state/sample, marks the lifecycle Faulted, and the
real v2 session returns a negative `DYNAMICS_RUN_FAULTED` response rather than
encoding poisoned values.
The rebuilt Windows Client was launched at
`src/AutoForge.Client/bin/Release/net10.0-windows10.0.26100.0/win-x64/AutoForge.Client.exe`
and operated through its local WebView2 session. Import of unknown profile,
revision and module references showed recoverable error notices without
changing the selected valid configuration. The short-deadline run visibly ended
Faulted at step 101 / 10.1 model seconds; the failure remained visible after a
dynamic diagnostic read and previously good telemetry rows were cleared. A
valid baseline run completed at step 3,361 / 336.1 model seconds. The active
runner was then paused and held at step 1 while waiting, resumed, stopped,
explicitly stepped, reset to a new step-0 session, read 17 `SYNTHETIC_DYNAMIC`
values at step 0, exported/re-imported the configuration, and closed. The
`/diagnostics`, `/vehicle-profiles`, `/execution-targets`, `/`, and unmatched
not-found routes loaded. The launched Client and its local DevTools listener
were closed and verified absent afterward.
The Step-04 worker report remains historical; the corrective evidence is in the
[Step-04A report](../prompts/Step-04A-Report.md). Step-04/04A/04B are accepted.
All model and telemetry values remain synthetic.
## Step-04B owner-drain corrective verification
Start and resume now reject before changing lifecycle state when a previous
advancement owner is still cleaning up after pause, stop or reset. The real
simulation page reports that rejection as an error rather than a successful
start/resume. Controlled barrier tests reproduced the prior defect: while the
cancelled owner was held in cleanup, Pause-to-Resume changed `Paused` to
`Running`, Stop-to-Start changed `Stopped` to `Running`, and Reset-to-Start
changed `Stopped` at step 0 to `Running`. After the guard, all three complete
status snapshots and sample counts remain unchanged on rejection; releasing
cleanup then permits exactly one retry and one new fixed step.
The final sequential Release build covered all nine solution projects and
succeeded with zero warnings and zero errors. The full solution test command
passed 126 unit and 81 integration tests (207 total; zero failed/skipped),
preserving all 202 Step-04A cases. Focused owner-drain regressions passed 4
unit and 2 integration tests. See the [Step-04B corrective report](../prompts/Step-04B-Report.md)
for exact commands and detailed contract evidence.
The actual native Windows Client smoke exercised start, active pause/resume,
stop/restart, reset, protocol-v2 telemetry read and close. It confirmed stable
paused/stopped steps, a reset session at step 0, `SYNTHETIC_DYNAMIC` telemetry,
and release of the closed session. This manual UI smoke does not claim to force
the cleanup interleaving; that exact race is established by the controlled
headless tests. The Client and its task-local WebView2 debug listener/profile
were stopped and cleaned up. Step-04/04A/04B are accepted; the prior numerical
thresholds and synthetic-only scope are unchanged.
## Step-05A live dashboard corrective verification
The corrected dashboard is shared by the actual `/diagnostics` static-v1 and
`/simulation` dynamics-v2 workflows. Focused headless checks passed:
- `dotnet test .\src\tests\AutoForge.UnitTests\AutoForge.UnitTests.csproj --configuration Release --no-restore --filter FullyQualifiedName~LiveDashboardGaugeRenderingTests`
— 7 passed, 0 failed, 0 skipped.
- `dotnet test .\src\tests\AutoForge.IntegrationTests\AutoForge.IntegrationTests.csproj --configuration Release --no-restore --filter "FullyQualifiedName~LiveDashboard|FullyQualifiedName~RealApplicationAndRenderedRoute"`
— 10 passed, 0 failed, 0 skipped. The selection includes controlled feed,
actual static/dynamic provider, rendered panel, lifecycle and startup
regression cases.
- `dotnet test .\src\tests\AutoForge.IntegrationTests\AutoForge.IntegrationTests.csproj --configuration Release --no-restore --filter FullyQualifiedName~RealApplicationAndRenderedRoute_RunAndReadDynamicTelemetryWithoutOracle`
— 1 passed, 0 failed, 0 skipped; the real `/simulation` route renders the
shared dashboard panel for an open dynamics session.
The required sequential full commands were then run:
```powershell
dotnet build .\src\AutoForge.slnx --configuration Release --no-restore
dotnet test .\src\AutoForge.slnx --configuration Release --no-build
```
The build succeeded for all nine projects, including the native Windows Client,
with 0 warnings and 0 errors. The MAUI XAML source-generation notice was
informational. The complete suites passed 133 unit and 91 integration tests
(224 total, 0 failed, 0 skipped), preserving the 207-case baseline and adding
17 dashboard-focused cases.
The actual Windows Client smoke used .NET SDK 10.0.402/runtime 10.0.13, MAUI
workload 10.0.403 / `maui-windows` manifest 10.0.110, Windows SDK target
10.0.26100.0 and WebView2 Runtime 156.0.4314.50. On `/diagnostics`, the worker
explicitly imported and selected the synthetic truck/profile/provider, opened
the static session, and started monitoring before selecting verified bindings.
The native editor visibly rejected an invalid range and an `rpm` signal bound
to `km/h` without adding a gauge or stopping the feed. A numeric gauge rendered
850 rpm; a bar first rendered 850 rpm below its configured range at 0% geometry,
then after editing to 0..1000 rendered 85%. Removing it and hiding the remaining
gauge did not stop polling. Stop/restart succeeded on the same session. Reset
created a new session with a stopped empty dashboard and no prior gauge values;
a fresh feed started on the replacement session, and close removed the panel.
On `/simulation`, the actual v2 session was paused at step 1 / 0.1 model seconds.
The dashboard rendered `SYNTHETIC_DYNAMIC` engine speed and vehicle speed with
logical request ticks, model step/time, configuration fingerprint and run ID.
After an explicit 200-step command, the readings updated to step 201 /
20.1 seconds (2,795 rpm and 19.7 m/s). Two later polls kept the paused model at
step 201. Reset opened a new step-0 session/run with an empty stopped dashboard;
the first new run sample displayed `dynamics-run-000002`, not the prior run.
Close released the panel. The existing execution-target, profile, home and
not-found routes loaded. Both temporary Client/WebView2 processes and their
isolated user-data profiles were stopped/removed; loopback debug listeners were
verified closed. See the [Step-05A corrective report](../prompts/Step-05A-Report.md)
for the named controlled fault/lifecycle evidence and full acceptance mapping.
After the final source-path cleanup, the rebuilt Client was launched again;
actual static-v1 and dynamics-v2 sessions each reached dashboard cycle 1 and
closed cleanly with the panel removed.
## Step-05B range and boundary verification
The Uncertain projection now computes its range status independently of source
quality. Before the fix, the real-provider regression failed on the actual
Uncertain -10 rpm sample: expected `OutOfRangeLow`, actual `null`. After the
fix, the focused unit suite passed 8 cases and the focused integration selection
passed 3 cases, including real-provider Uncertain -10/1100 readings rendered
through both numeric and bar gauges, Uncertain values at both exact bounds,
atomic invalid-definition vectors, and controlled Stop-drain cleanup.
The full post-fix Release build succeeded for all nine projects with 0 warnings
and 0 errors; the MAUI XAML source-generation notice was informational. The full
Release test suite passed 134 unit and 91 integration tests (225 total, 0
failed, 0 skipped).
On the exact rebuilt native Client and an isolated WebView2 profile, the static
editor configured a horizontal bar from the real `SYNTHETIC_STATIC` engine
signal and displayed 850 rpm. The dynamics editor configured a numeric gauge
from the real `SYNTHETIC_DYNAMIC` engine signal and displayed 260 rpm with
logical tick, model step/time, configuration fingerprint and run identity; the
model was paused at step 1 / 0.1 seconds. Both sessions closed with their
dashboards, the exact test processes and port 9351 were verified stopped, and
the isolated test profile was removed. The older shared WebView2 profile was
left untouched. These native observations confirm the dashboard baseline; the
Uncertain and malformed-input cases are established by controlled tests, not
claimed as native UI injection.
Step-05 with Step-05A and Step-05B was accepted by the orchestrator on
2026-10-29. These historical worker results remain distinct from that manual
acceptance.
## Step-06 display-unit verification
The shared shell defaults to Imperial (US) / Road and persists explicit
Metric/Imperial (US) and Road/Engineering selections through the native MAUI
Preferences adapter. Portable tests inject an isolated store; controlled read,
write, corrupt-value and invalid-selection cases exercise the same application
service without modifying platform preferences. Pure tests cover the specified
forward/inverse conversion vectors, affine temperature bounds, finite/overflow
handling, exact pass-through/unknown units and conversion before rounding.
Rendered real-provider coverage toggles all speed modes repeatedly on a paused
dynamics session and verifies unchanged source snapshots, gauge definitions,
trace/request count, configuration export, model state and bar percentage.
Current command outcomes and exact complete suite counts are recorded in the
[Step-06 worker report](../prompts/Step-06-Report.md). Native persistence and
draft-editor observations are reported separately from controlled storage
failure/error tests; headless verification does not substitute for the actual
Windows Client smoke. This section preserves the original Step-06 worker
evidence; the subsequent Step-06A correction and current review status are
recorded below. At that time, Step-07 recording had not yet run; replay remains
outside Step-07 scope.
The final sequential Release build completed all nine projects, including the
Windows Client, with 0 warnings and 0 errors. The full suite passed 178 unit and
92 integration tests (270 total; 0 failed/skipped), preserving the accepted
225-case Step-05B baseline and adding 44 unit plus one integration case. Focused
display-unit tests passed 44/44; the focused real static/dynamic/rendered
integration selection passed 3/3.
The rebuilt Windows Client was launched with task-local WebView2 debug port 9361
and an isolated temporary WebView profile. Its initial native shell showed the
Imperial (US) / Road first-use default before any session; the key had no saved
selection. A real static v1 session rendered 850 rpm unchanged after switching
the shared selectors. A real dynamics v2 session was paused at step 203 /
20.3 seconds and returned 20.1168 m/s. Its bar displayed 45 mph / 0..90 mph,
66 ft/s / 0..132 ft/s, 72 km/h / 0..144.84096 km/h, and 20 m/s /
0..40.2336 m/s; all modes kept 50% geometry, the exact 20.1168 m/s source
reading, and step 203.
The native range editor saved 0..90 mph as canonical 0..40.2336 m/s. An open
10..110 mph draft retained its mph labels and values after the global system
changed to Metric; saving produced source bounds
4.470400000000001..49.174400000000006 m/s and the equivalent Metric/Road
display range. An invalid 200..100 range surfaced an alert and preserved the
prior gauge; an unedited Save retained its exact existing source bounds.
Protocol-v2 telemetry displayed Imperial numeric readings while retaining
source `kg/s`, `N*m`, `degC` and exact source speed alongside them. Exported
configuration remained 2,789 characters, contained the unchanged SI target
20.1168 and no display-unit labels, and kept its fingerprint. Reset opened a
new stopped step-0 session/run with an empty dashboard; close released it.
The native app was closed and reopened: Imperial (US) / Road restored as a
saved choice. Existing diagnostics, execution-target, profile, home and
not-found routes loaded without an error banner. Only the exact temporary
WebView profile was removed; the preference created by this smoke was left at
the sensible Imperial/Road default, and no unrelated preferences were changed.
## Step-06A corrective verification
The original Step-06 review found invalid-selection initialization/storage
mutation and unavailable/wrong-type readout presentation gaps. Step-06A closes
those gaps without changing the approved defaults, mappings, editor policy,
model/source data or protocol. Step-06 remains **not accepted** pending manual
review; Step-07 had not yet run at that point. The final Step-07 evidence is
recorded below.
Before the fix, an invalid proposal on a fresh service loaded the saved
`metric|road` selection instead of leaving the untouched default; store reads
were not zero. The presentation regressions observed `"null"` for a null
numeric measurement, echoed `"not a number"` for a string-valued `m/s` signal,
and omitted an unconverted notice for unsupported-unit text. After the fix,
`PreferenceService_RejectsInvalidChoiceBeforeInitializingOrReadingStore`,
`RejectedSelectorEventReconcilesControlAndReadoutAndCanRetry` (both selectors),
`DiagnosticRoute_RendersNullWrongTypeUnknownAndUnsupportedStaticV1SignalsTruthfully`,
and `FormatSourceValue_PreservesJsonNullAndBooleanLexicalValues` pass as part
of the full suite. Real-provider controlled tests cover converted Uncertain,
Stale, Stopped and failed-display projections; protocol-v2 validation remains
strict.
The final sequential Release build completed all nine solution projects with
0 warnings and 0 errors; the MAUI XAML source-generation notice was
informational. The complete suites passed 179 unit and 105 integration tests
(284 total; 0 failed, 0 skipped), preserving the 270-case Step-06 baseline and
adding 14 corrective cases. Exact final commands:
- `dotnet build .\src\AutoForge.slnx --configuration Release --no-restore`
- `dotnet test .\src\AutoForge.slnx --configuration Release --no-build`
The rebuilt Windows Client smoke used the task-isolated WebView2 debug port
9361 and an isolated `LOCALAPPDATA`/WebView2 profile. First use showed
Imperial (US) / Road. A real static v1 session rendered 850 rpm unchanged. A
real paused v2 session at step 203 / 20.3 seconds retained exact source
20.1168 m/s while native readouts displayed 20.117 m/s, 72.420 km/h, 45.000
mph and 66.000 ft/s across all four selections. A native horizontal bar
displayed 72 km/h from source 20.1168 m/s over 0–144.84096000000002 km/h with
50% geometry. Its open Metric/Road editor draft remained in km/h after the
global selection changed to Imperial (US)/Engineering; closing and reopening
the app restored the task-isolated Metric/Road preference. Both static and
dynamics sessions closed and released their dashboards.
The native smoke did not inject a platform storage failure or malformed v1
values. Those cases are covered by controlled store tests and the real static
v1 rendered integration path, respectively. All native preference/profile
data was written under the task-owned isolated profile; that exact profile was
removed after both Client processes and port 9361 were closed. No unrelated
user preference/profile data was changed. See the [Step-06A corrective worker
report](../prompts/Step-06A-Report.md) for matrix-level evidence, limitations
and preservation review.
## Step-07 bounded telemetry recording verification
The final sequential Release build completed all nine solution projects,
including the Windows Client, with zero warnings and zero errors. The
MAUI XAML source-generation notice was informational. The full suite passed
179 unit and 124 integration tests (303 total; zero failed and zero skipped),
preserving the Step-06A baseline of 284 and adding 19 integration regressions.
The focused telemetry recording selection passed 19/19, including the
cross-scope capture-retention regression. Commands:
- `dotnet build .\src\AutoForge.slnx --configuration Release --no-restore`
- `dotnet test .\src\AutoForge.slnx --configuration Release --no-build`
- `dotnet test .\src\tests\AutoForge.IntegrationTests\AutoForge.IntegrationTests.csproj --configuration Release --no-restore --filter "FullyQualifiedName~TelemetryRecordingIntegrationTests"`
The rebuilt Windows Client was launched with isolated task-owned WebView2
profiles and local DevTools ports. A static v1 session captured seven actual
operations (including positive and no-response discovery, DTC and signal reads),
closed with `SourceClosed`, saved locally, then was loaded after app restart.
The readback showed `autoforge.synthetic-diagnostics v1`, endpoint `0x10`, raw
request/response payloads, and two decoded source signals (`850 rpm` and
`12.6 V`). A dynamics v2 session was paused at step 8 / 0.8 model seconds.
With gauges hidden, two actual `read.signals` observations captured exact
configuration/run and model step/time context. The panel was unmounted by
navigating Home; the session close stopped the capture as `SourceClosed`,
and its unsaved snapshot remained available on the dynamics route for explicit
save. After another Client restart, both archives loaded and exposed raw bytes
and decoded signals; the static capture had two signals and the dynamics
capture had five. The displayed Metric/Road selection remained unchanged.
That route-unmount smoke exposed a scoped-workspace lifetime bug; the workspace
was changed to application-wide singleton lifetime and
`CaptureWorkspace_IsSharedAcrossHostScopesAndRetainsClosedSourceCapture`
regresses it. Native post-fix evidence confirms the capture survives unmount
and source close. The UI also renders protocol versions and endpoint addresses
without Razor interpolation artifacts. Controlled tests, not the native smoke,
cover exact fixed limits, malformed archives, cancellation, collision and
filesystem failures.
The `LOCALAPPDATA` override isolated neither MAUI's
`FileSystem.AppDataDirectory` nor display preferences on this Windows
environment. The application initially listed zero archives; the two
task-generated captures were written to its app-local `recordings` directory,
then those exact IDs were removed after closing the Client. No pre-existing
capture was overwritten or deleted. The WebView2 profiles were task-owned
temporary directories and were removed. Metric/Road preferences were observed
but not changed. See the [Step-07 worker report](../prompts/Step-07-Report.md)
for frozen-matrix evidence, self-review and preservation details. Worker
completion remains pending manual orchestrator review; no Step-08 work began.
## Step-07A corrective recording verification
Step-07 remains historically **NOT ACCEPTED**; this correction preserves its
303-case baseline and native smoke evidence. Before the fixes, the focused
recording selection ran 23 tests with 17 passing and 6 failing. The failures
reproduced actual boundary defects: an empty `TransportOutcome.Response` was
rejected before it could be archived; source-accepted case-variant identities
were rejected on readback; a positive signal read with a null snapshot escaped
as `NullReferenceException`; a stopped capture remained in the session
registration list; an existing file used as the storage root listed as empty;
and an actual panel save collision was followed by catalogue refresh, which
replaced the visible save error with a malformed-archive error. The latter was
also rerun against the real panel event before correction and failed because
the rendered error no longer contained the collision message.
After correction, `TelemetryRecordingIntegrationTests` passed 26/26. The tests
exercise the actual v1/v2 observation and save/read paths, coherent raw-and-
decoded v2 mutation cases, deterministic stop/error/limit registration release,
invalid and missing storage roots, and the actual panel collision/error/retry
flow. The full Release build completed all nine solution projects, including
the Windows Client, with zero warnings and zero errors; the MAUI XAML source-
generation notice was informational. The full suite passed 179 unit and 131
integration tests (310 total, zero failed and zero skipped). Exact commands:
- `dotnet test .\src\tests\AutoForge.IntegrationTests\AutoForge.IntegrationTests.csproj --configuration Release --filter "FullyQualifiedName~TelemetryRecordingIntegrationTests"` (26/26)
- `dotnet build .\src\AutoForge.slnx --configuration Release --no-restore` (nine projects; 0 warnings, 0 errors)
- `dotnet test .\src\AutoForge.slnx --configuration Release --no-build` (179 unit + 131 integration; 310 total)
The rebuilt Windows Client was exercised through its local WebView2 DevTools
endpoint. Static v1 captured seven actual operations, then stopped/saved,
closed its source, restarted the app, and loaded/inspected the archive.
Dynamics v2 was paused at model step 7 / 0.7 seconds; with gauges hidden, it
captured four actual `read.signals` operations, then stopped/saved, closed its
source, restarted, and loaded/inspected the archive. Both readbacks exposed raw
bytes and decoded signals/context; Metric/Road preferences remained unchanged.
Native captures were `034e287ae82e4e14aa4d55e78581cbf1` (static) and
`69def5834cf742ce9fa729dd4bea1b16` (dynamics). Those exact task-generated
archive files and temporary WebView2/profile directories were removed after
the Client and DevTools listener were closed.
Native evidence verifies the normal record/stop/save/restart/load/source-close
workflow; it does not claim native filesystem-fault injection. Invalid roots,
save collisions, coherent malformed archives, and error/limit lifetime
cleanup are covered by controlled integration tests. The `LOCALAPPDATA`
override did not redirect MAUI `FileSystem.AppDataDirectory` in this
environment. See the [Step-07A corrective worker
report](../prompts/Step-07A-Report.md) for the frozen-matrix mapping,
pre/post-fix outputs, self-review and preservation review. Step-08 remains
blocked pending manual review; no Step-08 work began.
## Step-07B lock-safe session disposal verification
The pre-fix reproduction used the real diagnostic protocol/session, capture and
provider transport with a test-only subscription wrapper exposing controlled
dispose barriers. It held a capture callback at limit or observation-error
detach after the capture lock was released, started close/reset until its
second unsubscribe was entered, then released the callback. Both actual tasks
remained incomplete: `Operation completed=False; close/reset completed=False.`
The two focused cases failed as expected after their five-second safety
deadlines (two failed, zero passed); the bounded `dotnet test` host exited
normally after reporting both cases, so no deadlocked process tree required
termination. Command:
`dotnet test .\src\tests\AutoForge.IntegrationTests\AutoForge.IntegrationTests.csproj --configuration Release --filter "FullyQualifiedName~SessionCloseDuringCaptureDetach_DoesNotDeadlockOrLoseFailurePrefix"` (exit 1).
These are deterministic barrier-forced lock-cycle observations, not a
stress-timing claim.
The corrected session publishes one shared disposal completion and snapshots
the lifetime resources under the lifetime gate, then performs cancellation
and resource/control/transport disposal after releasing it. Concurrent callers
observe the same task and cleanup error; late registration rejects once close
ownership has been established. The focused set passed 29/29 recording tests,
including limit-close and error-reset, exact prefix/reason preservation,
replacement-session isolation, concurrent disposal and cleanup failure.
Commands and results:
- `dotnet test .\src\tests\AutoForge.IntegrationTests\AutoForge.IntegrationTests.csproj --configuration Release --filter "FullyQualifiedName~SessionCloseDuringCaptureDetach_DoesNotDeadlockOrLoseFailurePrefix|FullyQualifiedName~ConcurrentSessionDispose_SharesCompletionAndReportsCleanupFailure"` (3 passed)
- `dotnet test .\src\tests\AutoForge.IntegrationTests\AutoForge.IntegrationTests.csproj --configuration Release --filter "FullyQualifiedName~TelemetryRecordingIntegrationTests"` (29 passed)
- `dotnet build .\src\AutoForge.slnx --configuration Release --no-restore` (nine projects; 0 warnings, 0 errors)
- `dotnet test .\src\AutoForge.slnx --configuration Release --no-build` (179 unit + 134 integration; 313 total, zero failed/skipped)
The rebuilt Windows Client was rechecked through its local WebView2 DevTools
endpoint. Static v1 recorded two actual operations (DTC and signal reads),
stopped, closed its source while retaining the unsaved capture, then saved.
After app restart, the archive loaded with its raw bytes and decoded signals.
Dynamics v2 was explicitly stepped to model step 7 / 0.7 seconds, captured four
actual `read.signals` observations with gauges hidden, then stopped and closed
its source while retaining the unsaved capture before saving. A second app
restart loaded both static and dynamics archives with raw data and decoded
source/model context. Metric/Road preferences remained unchanged. The exact
task-owned capture IDs were `3b6bbb098a094ac2ba10c042dcd4fbf0` (static) and
`ac5b4d4e814d4905a0d454292b4dde87` (dynamics); their exact archive files and
task-owned temporary profiles were removed after closing the Client.
The controlled integration tests prove the close/reset lock-cycle resolution;
the native smoke verifies unchanged normal record/stop/close/save/restart/
readback behavior and does not claim native race injection. See the
[Step-07B worker report](../prompts/Step-07B-Report.md) for matrix mapping,
post-green source review and preservation checks. Step-07/07A remain NOT
ACCEPTED pending manual review; Step-08 remains blocked and was not started.