8.1 KiB
Recorded telemetry replay
Open a saved capture offline
Open Recorded telemetry replay (/telemetry-replay) from the shared
navigation. Refresh validates files in the Client's existing app-local
recordings directory. Select an explicit capture ID and choose Load saved
capture. Loading never auto-plays. No profile import, diagnostic session,
live feed, vehicle, model or configuration catalogue is required.
The current execution kind is RecordedReplay, with a fresh replay-instance identifier and the saved capture ID. Provider, target kind, profile revision, vehicle/session/configuration/run identifiers, protocol and completion are original archived metadata, not current vehicle evidence or verified identity. Raw/decoded inspection shows the last applied observation; before first and empty archives have an explicit no-observation/no-sample state. The source archive is never edited, saved again or rebound into the recording workspace by replay. Readback establishes consistency, not authenticity.
Incomplete and Error captures retain a prominent original state/reason/error. Before Play, Next, Previous or either seek, explicitly acknowledge that load's partial/error source. Stop, First, inspection and Close remain available. Acknowledgement survives navigation within the same load, but not a replacement load. Reaching the end means end of retained data, never a newly Complete source or recovery of unrecorded samples.
Two independent cursor values
- Position is integer
TimeSpanticks in[0, duration]. One second is 10,000,000 ticks. Duration is the last observation's capture elapsed time, or zero when empty; no trailing UTC/model interval is invented. - Applied count is the exact prefix length from zero to the retained count. Load, First and Stop mean before first: count zero, position zero and no current values. This differs from time seek zero if zero-time records exist.
Next applies exactly one observation. Previous rebuilds one fewer. Ordinal seek rebuilds exactly the requested count and sets position to that last record's capture time (zero for count zero). Time seek includes every record whose capture time is less than or equal to the requested tick, including all ties, but preserves the requested position between records. Manual navigation leaves Paused, or Completed when every record of a nonempty archive is applied. Empty seek zero remains Stopped; empty Next/Previous is visibly rejected.
Time-driven playback applies whole due-time groups in recorded sequence before publishing. All-zero archives complete without a timer; manual Next can still inspect a single record inside a tie. Completion depends on applied count, not position equalling duration. Completed requires explicit Restart or navigation, not an implicit replay when Play is pressed again.
Diagnostic logical ticks and model step/time remain per-sample source context. They do not determine ordering or duration: model resets and repeated model times remain exactly recorded. Modules are not represented as simultaneous. There is no interpolation, extrapolation, smoothing or model rerun.
Exact quality and last-Good reduction
Bindings are pre-indexed from actual positive archived signal snapshots using the first-observed module/signal/exact unit/origin rules. This index permits configuration before playback but contains no current measurements. Every navigation reconstructs the bounded prefix from scratch (maximum 4,096 observations), clearing current, last-Good and per-module status caches first. No later value can leak into an earlier prefix.
Only read.signals observations change signal state. A failed operation changes
only that module. A prior finite Good sample may be retained as visibly
recorded Stale, with its original observation/capture/diagnostic/model time
and failure reason. Uncertain is not last-Good authority. Without a prior Good,
a failed read is Error/--, never zero. Unknown, null/non-numeric, missing
signals and unit/origin changes use the same source-quality/range/last-Good
rules as the live dashboard. Unsupported origin changes
are also rejected at the existing strict archive boundary.
Example: A=850 Good at 1 s; B=12.6 Good at 2 s; A=1000 Uncertain at 2 s; A Timeout at 4 s. Seek 1.5 s yields only A=850 Good, with no B sample. Seek 2 s applies three records, A=1000 Uncertain and B=12.6 Good. Seek 4 s retains A=850 Stale from 1 s, not 1000, and leaves B Good. Previous restores the exact three-record Uncertain prefix; First clears both.
One cancellable owner and explicit controls
Play resumes Stopped/Paused. Restart is rewind-plus-Play and requires quiescence. Repeated active Play is idempotent and shares the owner. Pause freezes the position at the request instant, applies the due prefix, then cancels/drains the old wait. Resume rebases from that position, waiting only the remaining capture interval. Stop drains then clears to before-first while retaining the archive, definitions, speed and acknowledgement.
Speed defaults to 1x and accepts finite 0.25x through 1000x inclusive. Only capture elapsed intervals are scaled; source facts and results are unchanged. Pause and finish cleanup before changing speed, seeking, restarting, replacing or refreshing an archive. Rejected speed input rerenders to the committed value; invalid inputs are never clamped into success. Gauge edits remain atomic and may rerender during playback without advancing it; archive/drain operations reject overlapping edits.
The clock uses monotonic TimeProvider and a cancellable wait seam, decimal
overflow-safe tick scaling and individual waits bounded at 30 wall seconds.
This is not a hard real-time SLA or smooth-animation promise. A clock/wait
failure is visibly Faulted, preserving the last committed prefix/sample state;
the actual fault reason remains visible even after another command is rejected.
A non-cancellation failure during pending Pause/Stop cleanup also remains
Faulted rather than becoming a successful reset. Stop/First after cleanup
provides explicit recovery.
The factory returns a fresh session per page, not a disposed scoped singleton. The page drains loading/playback on Close or unmount. Generation/owner checks prevent delayed old loads/wakes from publishing. While an archive command or cleanup is pending, competing starts/navigation/replacements reject before mutation. Failed/cancelled replacements retain the previous validated context, definitions, values, rate and position, with a visible actual error rather than a healthy empty replacement.
Shared recorded gauges and display units
Use the shared in-memory editor to add zero through eight numeric/bar gauges. Exact source units and bounds remain canonical; only labels/readouts are projected to global display units. An open draft pins its selected display units until save/cancel/reset. Invalid edits are atomic, and unedited bounds retain their exact original source values. All gauges are labeled Recorded replay, even Good or Paused/Completed readings. Hide/show, remounting and global unit changes never start/stop/seek/read the player.
Only the existing global unit preference persists. Layout, definitions, cursor, acknowledgement and speed are runtime-only; a new load clears definitions/cursor/acknowledgement but keeps valid speed and global units.
Verification boundaries
TelemetryReplayTests covers real recorder/codec/temporary-filesystem input,
offline DI, independent exact-prefix fixtures, controlled monotonic waits and
barriers, typed atomic rejections, faults, replacement/unmount, source-byte
immutability, zero live/model calls and zero Save calls, plus actual rendered
route/editor events. Existing live-source regressions remain intact.
Native evidence uses the real Windows Client, app storage and saved static-v1/
dynamic-v2 captures; it is not controlled timing, race or fault proof.
Commands, counts, native observations and limitations are recorded separately
in the Step-08 worker report.
Step-07/07A/07B and Step-08 are accepted. Synthetic calibration authoring is a
separate feature and cannot change replay archives, cursor or playback ownership.