Two watch-stranding fixes (BULLETPROOFING.md M1, M4):
The exclusive-edit lock rode in the latest-wins context and was cleared
only by onDisappear, so an editor left open in a pocketed (or
force-quit) phone parked the watch's run indefinitely ("Editing on
iPhone..."). The scene-phase hook now publishes the locks as cleared
while the app is backgrounded - without forgetting them locally - and
re-asserts them on return to the foreground.
pushAll treated a failed updateApplicationContext as log-only, so a
payload past WatchConnectivity's size ceiling silently froze the watch
out of all future state. A failed push now retries with the
recently-completed tail dropped (display-only on the watch); only a
failure of the slim push too remains an error.
Claude-Session: https://claude.ai/code/session_01PVNBVKp5bcq52X722uMjwT
The HKWorkoutSession could only ever be born via the phone's one-shot
startWatchApp handoff (BULLETPROOFING.md H1-H3): a dropped handoff, a
watch crash/reboot, or a run engaged manually on the wrist left the
whole workout sessionless - no heart rate, no Health save, app
suspending wrist-down - and "End Current & Start New" swallowed the
handoff against the old session's idempotency guard.
Now the coordinator self-starts a session on any reconcile that finds
an active run with none running (SessionEndPlanner.shouldStart /
runToStart, activity type from the run's split), recover() re-adopts a
crash-orphaned session at launch, and a system-ended session salvages
its Health save instead of dropping it. A .finish decided for a
session younger than 30s demotes to .discard so the stale-context
races can't save junk workouts attributed to the wrong run; parallel
completions now pick the survivor deterministically.
Claude-Session: https://claude.ai/code/session_01PVNBVKp5bcq52X722uMjwT
DEVICE-COMMUNICATION.md: exhaustive reference for every phone<->watch
exchange - triggers, wire datapoints, transports (T1-T4), and the internal
structures each side mutates - plus end-to-end sequences and failure modes.
BULLETPROOFING.md: ranked gap analysis (H1-H3 session lifecycle, M1-M4
stranding/stale-UI traps, L-tier) with code evidence, failure scenarios,
fix directions, and a recommended fix order.
Claude-Session: https://claude.ai/code/session_01PVNBVKp5bcq52X722uMjwT
ExerciseListView is reached only via Settings > Splits > split > Edit >
Exercises, so its toolbar start button was effectively undiscoverable and
duplicated the home-screen split picker (the sole remaining start path).
Strip the button and its now-dead start flow: the active-workout prompt,
end-and-restart handling, started-workout navigation, and the AppServices
dependency. Also drop the changelog entry for the watch-launch wiring of
that button (f01e149) since the button no longer exists and never shipped.
Claude-Session: https://claude.ai/code/session_01PVNBVKp5bcq52X722uMjwT
A single "Cardio" library entry (running-in-place figure, duration-logged)
plus a matching Cardio starter split tagged .cardio (HealthKit .mixedCardio),
so the Apple Watch records the real aerobic workout while the phone logs the
time — no per-machine cardio entries or schema changes.
Claude-Session: https://claude.ai/code/session_012qw2itfzKyEJ1HpsFt8Ex4
Adds WorkoutActivityType.warmUp (HealthKit .preparationAndRecovery) and
.stretching (.flexibility), and retags the six starter splits that were all
mislabeled as Functional Strength:
- Warm-Up: Upper Body Warm-Up, Lower Body Warm-Up, Morning Wake-Up
- Stretching: Morning Mobility, Full Body Stretch, Evening Stretch
The split editor's activity picker surfaces them automatically (CaseIterable).
Older app versions decode the new raw values as the default type — additive and
not schema-gated, so no quarantine.
The "Start This Split" button on a split's exercise list minted and
saved the workout but never called WorkoutLauncher, so the Apple Watch
never came up when starting from there — only the home-screen split
picker launched it. Inject AppServices into ExerciseListView and call
launchWatchWorkout from start(), mirroring the picker path.
Claude-Session: https://claude.ai/code/session_01PVNBVKp5bcq52X722uMjwT
Premium/enhanced voices ship with the tier in their name (e.g. "Ava
(Premium)"), so appending "— Premium" produced "Ava (Premium) — Premium"
in the voice picker. Move the label logic to SpeechSettings.displayLabel
and skip the suffix when the name already contains the tier (still
name-only for default voices). Deterministic string helper, unit-tested.
Claude-Session: https://claude.ai/code/session_01BQcEWmAPA78338QuEwRkAh
Two per-split settings, with the global Settings values as defaults:
- restSeconds: Int? — per-split rest, used between sets and (in flow) between
exercises; nil falls back to the global default.
- autoAdvance: Bool? — flow mode: finishing an exercise rests, then opens the
next one hands-free, all the way through the split.
Both are optional, snapshotted onto WorkoutDocument at the start sites (no live
split link), and not schema-bumped — same degradation pattern as activityType.
A thin RunFlowView wrapper (iOS + watch) owns the on-screen log and swaps it via
.id(currentLogID) on hand-off, so the per-exercise ExerciseProgressView stays
per-logID and untouched; the between-exercise rest reuses the existing .rest
countdown as the terminal page. The mirror reuses the per-logID live channel:
the wrapper suppresses the boundary .ended teardown so it follows across
exercises, and ContentView re-keys the cover on frame.logID — no sync-bridge
changes.
Morning Wake-Up ships as a flowing 45s-work / 15s-rest routine.
New Rest & Pacing section in the split editor exposes both controls.
Speak exercise setup and form cues aloud with AVSpeechSynthesizer:
- Library detail: a speaker toolbar button reads the full reference aloud.
- Active workout: an opt-in "Speak Exercise Cues" setting speaks a brief
cue (Setup/Execution/Cues) when an exercise starts, hands-free.
- Settings › Voice: pick the voice (auto-prefers an installed enhanced/
premium English voice), a premium-download nudge shown only while on a
basic voice, and a Speed/Pitch/Volume sheet with Reset to Defaults.
On-device and offline; ducks other audio rather than stopping it. iPhone
only for now. Shared SpeechSettings is the single source of truth for the
voice/prosody, read fresh per utterance so changes preview live.
Claude-Session: https://claude.ai/code/session_01BQcEWmAPA78338QuEwRkAh
ingestFromWatch arbitrated by whole-document updatedAt, so concurrent edits
to the same workout on both devices (phone edits exercise A while the watch
completes exercise B) lost one side wholesale — the newer snapshot replaced
the other (H1).
Reconcile per log instead. WorkoutMergePlanner (pure, deterministic) unions
logs by id, resolves each by newest per-log updatedAt, and applies
phone-authored deletion tombstones so an absent log is never ambiguous
between "deleted on the phone" and "just added on the watch". Edits to
different exercises now commute — delivery order and offline gaps stop
mattering. A stale/duplicate push merges back to exactly the cached doc, so
ingest re-pushes authoritative state rather than writing.
The per-log updatedAt scaffolding shipped (unused) in schema v4; it's now
stamped by transition(to:) on status flips and a new touch() at the
content-only edit sites (order, notes, machine settings, adjusted entries,
new logs) on both phone and watch. deletedLogIDs is new: additive on the
wire and cache, phone-authored (deleteLog), pruned after a 30-day grace.
Because an older build rewriting a file would strip the tombstones and
resurrect a deleted exercise, WorkoutDocument schema bumps 4->5 (forward
gate quarantines old builds) and the cache bumps 5->6.
recomputeStatusFromLogs takes an injectable now: so the merge recomputes
status/end deterministically. WorkoutMergePlannerTests pins the decision
table (commute, per-log newer-wins, legacy-nil, watch-add, tombstone
honored/resurrect/union/prune, status recompute, no-op re-push);
WorkoutDocumentMapperTests gains the deletedLogIDs round-trip.
The HKWorkoutSession that keeps the watch app foregrounded was ended only
by ActiveWorkoutGateView's onChange(of: activeWorkouts) — a view-level side
effect. When a run ended from the phone while the watch app was
backgrounded (kept alive only by the session) or torn down and rebuilt with
an already-empty list, that onChange never fired: the session leaked (the
app kept re-foregrounding on every wrist raise) and finishAndSave() never
ran, so the HR/energy summary was neither saved to Health nor forwarded.
Move session-end off the view into a long-lived WorkoutSessionCoordinator
owned by WatchAppServices, driven by a new bridge.onWorkoutsChanged
callback fired after every authoritative cache mutation (phone push or the
watch's own optimistic edit). The decision is a pure SessionEndPlanner
seam (mirrors WatchCacheApplier): a running session ends only on a genuine
non-empty -> empty transition of the active set, so the launch race (session
running before the run doc syncs) resolves to .none and never discards a
run we haven't heard about yet. Same move as the live-mirror's
repairFromDurable, one layer down.
Watch-only; no schema or wire change. SessionEndPlannerTests pins the
decision table; the OS-initiated-end path (system ends the session itself)
stays a documented residual in PLAN-watch-session-end.md.
Sequence and state diagrams for every phone↔watch interaction (splits,
exercise drive, run end, edit locks, cold launch, session lifecycle),
companion to WATCH-SYNC.md's channel reference.
Claude-Session: https://claude.ai/code/session_01PVNBVKp5bcq52X722uMjwT
The rest/timed-work countdown deadline is shared by both devices, but the
page flip crossing it is a local ticker event — and stamping the *next*
phase's anchor at Date() when that event finally ran baked a sleeping
watch's lateness (throttled wrist-down ticker) into its next count-up,
leaving the two devices permanently offset with nothing on the wire to
correct.
Auto-advances now chain the anchor instead: the finished phase's computed
end (passed out of CountdownPhaseView) becomes the next page's PageAnchor,
with the next window derived from it via liveSnapshot(for:at:). A device
arbitrarily late to a boundary shows exactly what the on-time device shows,
and a stack of missed boundaries fast-forwards itself — each chained page
lands already-elapsed and advances on its own next tick, skipping the
start/stop haptics for boundaries that passed while asleep (only a
just-crossed boundary buzzes).
The remoteAnchor* fields are generalized into one PageAnchor (remote frames
and chained auto-advances are the same concept: a page whose timer counts
from a known instant); the phone's Live Activity emit honors it unchanged.
Live frames ride sendMessage, which is reachable-only — and phone→watch
reachability drops exactly when the user is swiping on the phone (wrist
down). A frame that failed to send was staged but never retried until a
reachability edge, and a frame lost outright desynced the run until the
next human transition — sometimes forever, since a reconnect could even
re-send the stale staged frame and yank the peer backwards.
Four fixes, symmetric on both bridges and both run screens:
- A send that fails while nominally reachable now retries with a short
backoff (a few times per staged message) instead of being swallowed.
- Receiving a frame that outranks the staged outbound one drops the
staged frame, so a reconnect re-send can't move the run backwards;
a delivery the staged frame outranks is ignored as stale.
- Every staleness comparison now tie-breaks the shared version sequence
on the frame's wall-clock anchor (LiveProgress.isNewer) — after a
lost frame both devices can mint the same version, and the later
human action must win.
- Durable repair: when the absorbed workout doc shows sets completed
beyond anything the open run screen recorded or followed, it jumps
forward to the first unfinished set's work page — a lost frame now
degrades to a briefly-stale page instead of a stuck one.
The half-and-half split (paged timer over the looping form-guide figure)
is now iPhone-only. On the watch the ExerciseProgressView TabView fills the
whole screen — the figure slot is gone from both the active flow and the
Completed state, and the ExerciseFigure sources and ExerciseMotions
resources are dropped from the watch target since nothing there uses them.
Claude-Session: https://claude.ai/code/session_01HJDQQDA9QdP8zByg43H5v3
The orbit was always a real camera orbit — figure and props share one
rigid rotation — but a bar's screen-space angle authored the wrong 3D
rod: the default horizontal encoded a rod along the body axis, so
barbells hovered fixed on screen and vanished at the head-on view
where they should span widest. Line props now take "axis": "z" (both
renderers in lockstep, fixture-pinned): the world left-right direction
projects through the camera pitch like the floor quad — end-on plates
in profile, full span face-on, swinging with the hands in between.
Applied to the ten cross-body bars; vertical handles were already
orbit-invariant.
Goblet Squat's hand pins sat so close to the shoulders that the
two-bone IK was degenerate, flipping between a chicken-wing and an
elbow-behind solve; re-pinned level with the shoulders so the elbows
tuck straight down through the whole rep.
Claude-Session: https://claude.ai/code/session_01HJDQQDA9QdP8zByg43H5v3
The library grows to 64 exercises: arm circles, torso twist, leg
swings, hip and neck mobility, marching, calf raises, and a full
stretching set (forward fold, quad, calf, chest, triceps, hip flexor,
butterfly, cobra, child's pose), each with an authored motion rig and
reference page. Eight new starter splits join the catalog — Upper and
Lower Body Warm-Up, Morning Wake-Up, Morning Mobility, Full Body
Stretch, Evening Stretch, Free Weight Basics, and Full Body Machines —
regenerated deterministically at split schema v3 with fixed ULIDs.
Claude-Session: https://claude.ai/code/session_01HJDQQDA9QdP8zByg43H5v3
Every completed set now writes a SetEntry (reps/weight or seconds),
pre-filled from the plan by transition(to:) so the list checkbox, both
run flows, and One More all capture for free; reset clears, skip keeps
partials. The rest and finish pages show the just-done set as a pill
that opens a stepper sheet for correcting reps and weight (2.5 lb /
1.25 kg steps). The Weight Progression chart plots the top-set actual
weight and workout volume sums recorded sets, both falling back to the
plan for legacy logs via effectiveSetEntries.
Storage side of UX #3 rides along: plan weights are Double now.
Schema bumps: SplitDocument 2→3, WorkoutDocument 3→4 (a fractional
weight fails an older Int decode, and a rewrite would strip the
irreplaceable actuals), SwiftData cache 4→5. A per-log updatedAt is
reserved for the future cross-device log merge.
Claude-Session: https://claude.ai/code/session_01HJDQQDA9QdP8zByg43H5v3
iCloud Drive writes now flow through a persistent WriteBacklog sidecar
(drained with backoff, flushed on backgrounding, wiped with the cache on
account change), so a save can never be lost to a transient coordinator
error. A status banner on the workout list surfaces stuck syncing.
Also: the split picker gains a Recent section with day labels, split
rows fold SplitItem into SplitListView, and list rows dim the multiply
sign in sets-by-reps.
Claude-Session: https://claude.ai/code/session_01HJDQQDA9QdP8zByg43H5v3
Changelog entries for the Live Activity, mid-workout library adds,
backups, diagnostics, watch-only Health recording, navigation and
accessibility improvements, and the bug fixes; README key features
updated to match.
Claude-Session: https://claude.ai/code/session_01HJDQQDA9QdP8zByg43H5v3
A new Settings > Diagnostics screen reports the ubiquity container and
account status, per-document download and eviction state, network
readiness, and a count of documents skipped by the schema-version
forward gate — surfaced to help debug why a file isn't syncing.
Claude-Session: https://claude.ai/code/session_01HJDQQDA9QdP8zByg43H5v3
Adds a local ZIP backup/restore of the iCloud document tree via the
IndieBackup package, surfaced in Settings with retention controls. A
restore suspends the sync observer, mirrors the files, then rebuilds the
SwiftData cache; opening a shared .workoutsbackup file restores it. The
engine exposes the container Documents root and a restore lifecycle
(isRestoring guards a concurrent connect), and the backup file type is
registered for open-in-place.
Claude-Session: https://claude.ai/code/session_01HJDQQDA9QdP8zByg43H5v3
A new iOS widget extension shows the active exercise, its phase, and the
work/rest countdown on the lock screen and in the Dynamic Island, driven
by the run flow's live frames so locking the phone mid-set keeps the
timer. The activity is seeded on open, refreshed on every page settle,
dismissed when the flow is left, and cleared on next launch if stranded.
Unifies the build-info stamping across all targets via a YAML anchor.
Claude-Session: https://claude.ai/code/session_01HJDQQDA9QdP8zByg43H5v3
Starting a workout — from the split picker or a split's exercise list —
now drops straight into its log screen once the cache catches up, via a
shared StartedWorkoutNavigator. Adds VoiceOver labels/values to the log
checkboxes and the settings button, a color-independent numbered legend
and spoken summary to the heart-rate-zone bar, and Dynamic Type scaling
to the run-flow badges and timer.
Claude-Session: https://claude.ai/code/session_01HJDQQDA9QdP8zByg43H5v3
New Workouts Watch AppTests bundle wired into the watch scheme. Extracts
the phone-to-watch cache apply/prune into a pure, session-free
WatchCacheApplier seam and makes the HR-zone bucketing a nonisolated
static, so both can be unit-tested off the main actor without a live
WatchConnectivity session.
Claude-Session: https://claude.ai/code/session_01HJDQQDA9QdP8zByg43H5v3
The phone no longer writes estimated Health workouts: the watch, which
runs the live session, is the sole recorder. Replaces WorkoutHealthWriter
with a WorkoutHealthDeleter that only removes a legacy phone-estimate
workout when its record is deleted here, drops the MET calorie table and
the phone's write/read Health scopes, and keeps phoneEstimate decodable
for existing documents.
Claude-Session: https://claude.ai/code/session_01HJDQQDA9QdP8zByg43H5v3
Editing a workout's start date now removes the file at its old month
bucket so the record no longer duplicates on the next reconcile. Seed
reconcile re-checks the tombstone veto before overwriting an upgraded
seed. The watch applies authoritative-empty pushes so remote deletes
prune, and a re-saved finished workout keeps its original end time.
Adds unit tests for the mappers, path bucketing, and status machine.
Claude-Session: https://claude.ai/code/session_01HJDQQDA9QdP8zByg43H5v3
Pick any exercise from the full library while a workout is running, not
just the ones in its split. The new exercise's plan is seeded from the
most recent log of that exercise, else the library's authored Defaults
line, else a plain 3x10. Adds a searchable two-section picker sheet and
a **Defaults:** bullet to all 47 library reference pages.
Claude-Session: https://claude.ai/code/session_01HJDQQDA9QdP8zByg43H5v3
Weight moves up beside the exercise name (matching font), sets × reps
sits below it with a dimmed ×, and recorded machine settings show as a
secondary line under the name — no placeholder when unconfigured.
Single-line bodyweight rows center against the checkbox.
Claude-Session: https://claude.ai/code/session_01P152LxjZ4vePHsSorQa5Jf
Abdominal's pinned hands used elbow -40 as the IK plane hint, drawing
the arms hyperextended (user-reported). Flipping the hint bends the
elbows the natural way while the hands stay on the handles. Same class
of fix for the milder cases: Arm Curl and Shoulder Press elbows and
Calfs knees clamp to -8, Side Plank's raised arm to the -70 ROM cap.
The whole library now passes render.py --strict with zero warnings,
making it a valid verification gate. Fixtures regenerated; 48 tests
green.
Claude-Session: https://claude.ai/code/session_01HJDQQDA9QdP8zByg43H5v3
render.py --export bundles the 25 new motion rigs and info pages;
figure-fixtures.json regenerated for all 47 (the Swift solver reproduces
every new motion, including negative-y pins and the new prop uses). The
fixture count assertion moves to 47, and the no-bundled-motion test now
uses Treadmill - a permanent COVERAGE.md exclusion - since Bench Press
exists. 48 tests green; watch target builds. Coverage handoff complete -
TODO-coverage.md retired.
Claude-Session: https://claude.ai/code/session_01HJDQQDA9QdP8zByg43H5v3
Pec Deck and Rear Delt Fly author the face-on and from-behind cameras;
Triceps Pushdown and Face Pull are the cable station's two entries per
the coverage model (cable curl/crossover are considered exclusions).
Claude-Session: https://claude.ai/code/session_01HJDQQDA9QdP8zByg43H5v3
The biggest hole in the coverage matrix - the hinge had zero entries in
any modality. Deadlift and Romanian Deadlift are the library's first
standing (camera.zoom) and first barbell motions.
Claude-Session: https://claude.ai/code/session_01HJDQQDA9QdP8zByg43H5v3
The rig is ~211 canvas units tall standing but the canvas has 152 above
the ground line - every motion to date is seated or on the floor. A
per-motion "camera": {"zoom": ...} now scales the drawn output (figure,
props, mat, stroke widths) about the ground-center anchor in both the
reference renderer and the in-app view. Pure view transform: pins, prop
coordinates, and the Swift-solver fixtures stay in full-size authored
units; zoom 1 is byte-identical to before.
Claude-Session: https://claude.ai/code/session_01HJDQQDA9QdP8zByg43H5v3
With the equipment layer carrying world-space 3D form, the prop-free
gate comes off: machines now get the same slow orbit as the bodyweight
moves, their seats, cables, bars, and rollers turning with the figure.
Closes out the orbit-for-all-exercises plan.
Claude-Session: https://claude.ai/code/session_01HJDQQDA9QdP8zByg43H5v3
Hollow Body Hold, Leg Raises, and Reverse Crunch hid the far arm
because it was never authored - a leftover from the planar rig. Both
arms are now posed, so the far arm reads as the standard light member
behind the near one, the same visual language every other exercise
uses, and the figures stay truthful from any viewpoint. With no
remaining users, the hide mechanism is deleted from both renderers,
the motion schema, and the docs.
Claude-Session: https://claude.ai/code/session_01HJDQQDA9QdP8zByg43H5v3
Scene shapes, cable anchors, bar angles, pad perpendiculars, and roller
offsets all resolve in the authored view exactly as before, then rotate
about the world-vertical axis through the root anchor - the same
resolve-then-rotate pattern as the figure's pins and the mat - so at the
authored yaw every exercise renders bit-identically to today, and under
an orbiting camera the equipment turns with the figure while staying
welded to its hands and feet. Scene lines gain an optional depth plane
(z) and slab extrusion (depth) so seats, backrests, and platforms keep
form edge-on; the rect shape is retired (re-authored as slab lines).
All 14 machines' props re-authored with depths and verified at eight
orbit angles. The fixture snapshots move into the pipeline as
render.py --fixtures and now cover orbit-presentation samples with
resolved prop primitives for a spread of prop flavors; the in-app
renderer resolves props in MotionSolver (lockstep with resolve_props)
and the view just draws primitives.
Claude-Session: https://claude.ai/code/session_01HJDQQDA9QdP8zByg43H5v3
The floor rectangle was screen-locked, which broke the illusion the
moment the camera orbited. It is now a world-space quad on the ground
plane, sized to each motion's projected footprint across its key frames
and rotated through the same camera as the figure - a long rectangle in
profile, a parallelogram mid-orbit, end-on when face-on. Both renderers
in lockstep; fixtures unaffected (the mat is a pure addition).
Claude-Session: https://claude.ai/code/session_01LEoff8bXGBS83tK1c55Mf7
The default camera pitches down 10 degrees, so the floor reads as a
plane (drawn as a rectangle) and near/far contacts straddle it.
Elevation is pure presentation - IK pins solve in the flat authored
view and the posed body tilts, the same pattern as the orbit, so
authored canvas targets never go out of reach. The leg-extension
roller moves up onto the shin above the ankle and the leg-curl roller
tucks under the heel. Fixtures and reference test values regenerated
for the pitched camera.
Claude-Session: https://claude.ai/code/session_01LEoff8bXGBS83tK1c55Mf7
Shoulder and pelvis widths grow to human-like proportions per profile
(shoulders wider than hips for neutral/male, reversed for female) and
are now drawn — bars across the attach points that read near-full-width
face-on and as a shoulder/hip nub in profile, so limbs visibly hang
from a torso instead of a point. Orbiting no longer re-solves IK pins
in the rotated view (pins are canvas targets in the authored camera):
the pose resolves first and the posed body rotates, which fixes hands
sticking to stale screen points mid-orbit (Cat-Cow, Bird Dog, Plank).
Leg Extension and Leg Curl swap their ankle bars for a machine roller
disc — a new `roller` prop riding the shin's press side. Fixtures
regenerated; both renderers updated in lockstep.
Claude-Session: https://claude.ai/code/session_01LEoff8bXGBS83tK1c55Mf7
Bodyweight motions (no equipment layer, no hide list) now turn the
camera a full revolution every 24 seconds while the motion loops, so
the exercise reads from every side; machine exercises keep their
authored view, since scene equipment is a view-locked billboard and a
hide list describes a single viewpoint.
Claude-Session: https://claude.ai/code/session_01LEoff8bXGBS83tK1c55Mf7
The library's planar world-angle rig becomes a genuine 3D anatomical
model: skeleton.json holds bone-length profiles (real shoulder/pelvis
widths, feet, neutral/female/male) and per-joint ROM; motions pose
joints with anatomical angles (flexion/abduction/rotation from neutral
standing) under a per-exercise orthographic camera, resolved by
kinematics.py (3D FK, analytic two-bone IK with anatomical write-back)
and validated against physiological ranges. All 20 sagittal motions
were migrated by planar decomposition with 0.00 px golden parity against
the old renderer — relabeled to true anatomy, since shading is now
near-dark/far-light by camera depth rather than by limb suffix — and
the face-on machines are re-authored honestly: Abductor/Adductor with
real hip abduction (the foreshortened "frontal" profile is retired) and
Rotary with genuine spine axial rotation. Figures gain articulated
feet; profiles swap without touching a single motion script; --orbit
sweeps the camera 360° while a motion loops.
The in-app SwiftUI renderer (iOS + watch) is ported to the same model
and consumes the exported motions verbatim; figure-fixtures.json pins
its geometry to the Python pipeline within 0.5 px across every
exercise, key frame, tween, and orbit sample. Also makes the watch
bridge logger nonisolated for the newer SDK's stricter isolation
checking.
Claude-Session: https://claude.ai/code/session_01LEoff8bXGBS83tK1c55Mf7
rzen