Feat/genui - #64

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In @.github/workflows/release.yml:
- Around line 49-51: Update the linker-patching step after flutter pub get to
require and quote PUB_CACHE, restrict find to resolved jni-*/src/CMakeLists.txt
targets, and make the replacement idempotent so existing --build-id=none flags
are not duplicated. Remove the failure suppression and ensure the step exits
nonzero when no target file is found or patching fails.

In `@workout-logger/lib/genui/src/a2ui_panels.dart`:
- Around line 46-65: Update the trailing label Text in the Row alongside the
title to be wrapped with Flexible, and configure it with maxLines: 1 and
TextOverflow.ellipsis so model-provided labels cannot cause a RenderFlex
overflow.

In `@workout-logger/lib/genui/src/a2ui_parser.dart`:
- Around line 208-237: Update _extractJson to first attempt jsonDecode on the
complete stripped text and immediately return the decoded Map or List when
successful. Only run the existing balanced-candidate scan when whole-text
decoding fails, preserving the current fallback behavior for prose-wrapped JSON.
- Around line 78-98: Update the effective-props construction in the parser
around A2UiProps.stringKeyed and _parseChildren so that when json['props'] is a
Map, literal child-related keys from the outer json object are copied into
effective only when those keys are absent from props. Preserve props values when
both levels define the same key, and keep the existing flat-payload behavior
unchanged.

In `@workout-logger/lib/genui/src/a2ui_renderer.dart`:
- Around line 26-35: Update A2UiRenderer.build around the specFor lookup to emit
a debug-only diagnostic containing node.name when spec is null, then preserve
the existing SizedBox.shrink() fallback. Use the project’s existing debug
logging mechanism rather than changing rendering behavior.

In `@workout-logger/lib/genui/src/a2ui_spec.dart`:
- Around line 54-58: Update the A2UiSpec contract methods buildWidget and render
to use named parameters for all three arguments, then update every
implementation and call site consistently, including the renderer and parseProps
flow. Preserve the existing behavior and argument types, or document an explicit
exemption if the Widget build-style API must remain positional.

In `@workout-logger/lib/genui/src/a2ui_theme.dart`:
- Around line 41-42: Guard the public A2UiTheme palette contract so seriesColor
never evaluates a modulo operation with an empty seriesPalette. Update the
A2UiTheme constructor to assert or otherwise enforce a non-empty palette, while
preserving the existing indexed cycling behavior for valid palettes.

In `@workout-logger/lib/genui/src/components/dynamic_chart.dart`:
- Around line 223-249: Update _pie to filter props.series.first.values to
positive entries before calculating total or creating PieChartSectionData,
preserving each entry’s original index so props.labels remains aligned. Return
A2UiEmptyPanel when no positive values remain, and use the filtered values for
percentages and section geometry.

In `@workout-logger/lib/genui/src/components/metric_gauge.dart`:
- Around line 217-221: Update _GaugeArcPainter.shouldRepaint to also compare the
track property from oldDelegate, ensuring the painter repaints when the
background arc color changes while preserving the existing progress, from, and
to comparisons.

In `@workout-logger/lib/genui/src/components/stat_card.dart`:
- Around line 85-94: Update the value-formatting logic around rawValue, unit,
and the contains check to determine whether the unit is already present only
when it matches the trimmed suffix of rawValue. Preserve the existing fallback,
empty-unit handling, and spacing behavior while preventing short units from
matching unrelated text.

In `@workout-logger/lib/main.dart`:
- Around line 141-143: Update the CoachToolService constructor and every call
site, including the shown dependency injection call, to use named parameters for
all three dependencies. Preserve the existing dependency wiring while making
each argument explicit by name.

In `@workout-logger/lib/models/models.dart`:
- Around line 138-151: The calculateVolume method should use final for locals
that are not reassigned: replace the mutable effW declaration and conditional
assignments with a single final conditional expression, and change the drops
iteration variable from var drop to final drop.
- Around line 138-157: Persist assisted-bodyweight semantics so default volume
calculations use effective load rather than assistance weight: update
WorkoutSet.calculateVolume and its persisted fields in
workout-logger/lib/models/models.dart:138-157, aggregate the persisted
assisted-set calculation in ExerciseLog.totalVolume at
workout-logger/lib/models/models.dart:240-243, and update the logging flow at
workout-logger/lib/screens/workout_flow_screen.dart:548-555 to store assistance,
extra load, and a body-weight snapshot (or the effective load) for assisted
exercises.

In `@workout-logger/lib/screens/widgets/exercise_input_section.dart`:
- Around line 75-76: Update the assisted-load display calculations around
isAssistedBW and the equivalent values at the referenced later section to pass
every displayed weight through settings.toDisplay, including effectiveWeight,
settings.userBodyWeight, and currentWeight, while keeping the stored kilogram
values unchanged.
- Around line 75-76: The assisted-exercise classification is inconsistent
between the load panel and _InputRow for dips and push_ups. Reuse the existing
isAssistedBW result when constructing _InputRow, pass it into the widget, and
update _InputRow to use that value for its labeling instead of maintaining a
separate exercise-ID predicate.
- Around line 194-223: The handle selector in build must not display an
unpersisted first handle as selected; require or persist an explicit selection
before logging. In
workout-logger/lib/screens/widgets/exercise_input_section.dart:194-223, update
the active-selection logic accordingly. In
workout-logger/lib/services/workout_provider.dart:506-515, update
setExerciseHandle so existing recorded set handles are never rewritten; lock the
handle after the first set or create a separate log for each variation.

In `@workout-logger/lib/screens/workout_flow_screen.dart`:
- Around line 322-325: Make _loadLastSessionData pass the current log handle
from the onHandleChanged callback. In
workout-logger/lib/services/workout_provider.dart lines 674-696, require an
exact handle match whenever handle is non-empty, excluding logs with null or
different handles. Apply the same rule in lines 699-713 for last-session lookup,
using any legacy fallback only after no exact match exists.

In `@workout-logger/lib/services/ai/coach_tool_service.dart`:
- Around line 414-425: Clamp the model-provided days value in the tool method
before the loop that calls hh.sleepNight, reusing the existing _limitArg helper
and its supported range as other tools do. Keep the default of 14 for missing
input, then iterate using the clamped value.
- Around line 523-546: Update
workout-logger/lib/services/ai/coach_tool_service.dart#L523-L546 so
_getHealthMetrics uses the requested days window rather than always querying one
week, and align the get_health_metrics declaration with the sleep fields
actually returned instead of promising resting HR or readiness data. Update
workout-logger/lib/services/ai/coach_tool_service.dart#L599-L606 by removing the
resting_hr x_metric branch/description unless it is backed by a real data
source; no additional metric should be advertised without implementation
support.
- Around line 625-642: Remove the synthetic sleep fallback block guarded by
xVals.length < 2 in the analysis method, including the generated synthSleep
values and additions to xVals, yVals, and points. When fewer than two real
paired points remain, return the existing insufficient-data error so
correlation, regression, and chart points are never produced from fabricated
health data.
- Around line 714-720: Update get_muscle_group_volume’s matching logic to
resolve each requested group with _resolveMuscleGroup, then compare resolved ids
using _wp.getMuscleGroupName(e.primaryMuscle) instead of raw display-name
substring matching. Preserve valid group requests such as “Quadriceps” and
“Lower Back”, and include secondary muscle activations in the aggregation
alongside each exercise’s primary group.
- Around line 26-28: Update CoachToolService’s constructor to accept the
optional HealthHistoryManager as a named parameter, then update all
instantiations to pass it using healthHistory: while preserving existing
dependency behavior.

In `@workout-logger/lib/services/ai/gemini_ai_service.dart`:
- Around line 91-95: Restrict daily-quota classification in
_isDailyQuotaExhausted within
workout-logger/lib/services/ai/gemini_ai_service.dart:91-95 and
is_daily_quota_exhausted in workout-logger/scripts/test_gemini_api.py:63-64 to
daily-limit identifiers such as GenerateRequestsPerDay or free_tier_requests, or
exact daily metrics; remove generic QuotaExceeded and RESOURCE_EXHAUSTED matches
so minute-scale rate limits continue through retry-delay handling.
- Line 496: Preserve each function-call ID through the response flow: in
workout-logger/lib/services/ai/gemini_ai_service.dart lines 457-496, retain
fc['id'] on each FunctionCall and include the matching ID in every emitted
functionResponse; in workout-logger/scripts/test_gemini_api.py lines 242-255,
copy fc["id"] into each generated functionResponse.
- Around line 98-105: Make the Gemini fallback payload compatible with each
selected model: update _getFallbackModel and the request-building logic around
thinkingConfig so gemini-2.5-flash uses thinkingBudget rather than Gemini 3.x
thinkingLevel, or remove that fallback. Apply the same compatible fallback chain
and per-model thinkingConfig normalization in
workout-logger/lib/services/ai/gemini_ai_service.dart at lines 98-105 and
258-259, and workout-logger/scripts/test_gemini_api.py at lines 67-72 and 221 so
retrying with a new model rebuilds the configuration.

In `@workout-logger/lib/services/interfaces/ml_service_interface.dart`:
- Around line 78-83: Update the documentation for MLService.recommendSets to
state that pastSessions must be ordered most recent first, with index 0 as the
latest prior session and index 1 as the preceding session. Replace the “past 3
sessions” wording with documentation matching the implementation’s two-entry
usage.

In `@workout-logger/lib/services/ml_service.dart`:
- Around line 449-457: Update the post-deload recovery branch in the
recommendation logic to avoid embedding raw set.weight with a hardcoded kg label
in the reasoning string. Either remove the weight value from this message or
route its display through the existing presentation-layer formatting, such as
SettingsProvider.formatWeight, so both units and numeric formatting respect the
user’s settings.
- Around line 375-391: Update deload detection in recommendSets to require the
comparison session to be recent, preventing permanent resets from triggering
recovery from an older workout. Use effective load, consistent with set.volume,
when calculating w0 and w1 for assisted bodyweight exercises instead of raw
set.weight, and cap any recovery recommendations derived from refSets near the
last performed weight rather than allowing high-confidence loads far above
lastSession.

In `@workout-logger/lib/services/settings_provider.dart`:
- Around line 51-53: The body-weight validation in the storage-load path and
setUserBodyWeight must reject non-finite or non-positive values. Update both
paths to accept only finite values greater than zero, persist only valid inputs,
and use 70.0 when the stored userBodyWeight is invalid.

In `@workout-logger/lib/services/workout_provider.dart`:
- Around line 660-671: Update getRecommendations so handle-scoped requests do
not pass the exercise-wide _growthModels[exerciseId] into recommendSets. Key
growth models by both exerciseId and handle when the training contract supports
it; otherwise pass no growth model for handle-scoped recommendations while
preserving the existing exercise-wide behavior for unscoped requests.

In `@workout-logger/lib/theme/a2ui_app_theme.dart`:
- Around line 1-2: Update the imports in the theme adapter to use the public
genui barrel exported by lib/genui/a2ui.dart, importing A2UiTheme through
package:repforge/genui/a2ui.dart instead of the internal src/a2ui_theme.dart
path.

In `@workout-logger/scripts/test_gemini_api.py`:
- Around line 93-98: Update the retry/fallback flow in the test_gemini_api
function so a quota response on the final attempt cannot fall through and return
None after selecting a fallback. Track fallback attempts independently from the
normal retry limit, or raise a clear error when no attempts remain; ensure the
function always returns its expected dict result or an explicit exception before
callers invoke res1.get(...).

In `@workout-logger/test/genui/a2ui_prompt_test.dart`:
- Around line 39-42: Update the example extraction around start and end so the
closing brace search is limited to the region beginning at markerIndex, rather
than using section.lastIndexOf('}') across the entire section. Preserve the
existing substring and validation behavior while ensuring later prompt content
cannot extend the extracted example.

In `@workout-logger/test/genui/a2ui_purity_test.dart`:
- Around line 9-11: Update the _forbiddenPathPattern in the purity test to
include the app-specific data directory alongside theme, models, services, and
screens, ensuring package and relative imports from data are rejected by the
existing guard.
- Around line 95-96: Hoist the cast-matching RegExp used in the loop in the
relevant purity test function into a top-level _castPattern declaration beside
_forbiddenPathPattern and the other shared patterns. Replace the inline RegExp
construction in the per-line check with this shared pattern, preserving the
existing matching expression and behavior.

In `@workout-logger/test/genui/a2ui_registry_test.dart`:
- Around line 133-136: Update the A2UiNode instantiation in the test to use the
const constructor, preserving its existing name and props arguments so
prefer_const_constructors is satisfied.

In `@workout-logger/test/genui/a2ui_renderer_test.dart`:
- Around line 153-158: Update the comment near the GridContainer assertions to
remove the contradictory statement that the node is dropped entirely; document
only that excluding items yields a real zero-children node and the expected
non-crashing behavior, matching the assertions.

In `@workout-logger/test/genui/a2ui_robustness_test.dart`:
- Around line 144-145: Strengthen the minY assertion in the chart axis test to
require the axis minimum to bracket the dataset’s true minimum of -50, rather
than merely being below -10. Keep the existing maxY assertion and the rest of
the chart test unchanged.
- Around line 80-84: Update the test names in both the parser loop and the
“renderer never throws” loop to derive from each payload’s truncated string
instead of the mutable list index. Preserve enough payload content for stable
failure attribution while keeping names concise.

In `@workout-logger/test/genui/a2ui_theme_test.dart`:
- Around line 138-143: Update the Container finders in both assertions around
the panel decoration checks to scope them to the panel widget under test instead
of using the global find.byType(Container). Apply the same scoped finder change
at the second assertion near line 159, preserving the existing padding, color,
and border expectations.

In `@workout-logger/test/genui/components/dynamic_chart_test.dart`:
- Around line 152-161: Add a widget test alongside the existing pie-chart test
that supplies mixed-sign values such as [60, -40], verifies the pie chart
renders without throwing, and confirms the expected fallback behavior from _pie
when percentage calculation receives invalid values.
- Around line 163-175: Extend the testWidgets case for multi-series non-pie
charts to also pump a single-series chart and a multi-series pie chart,
asserting that A2UiLegend is absent in both cases while preserving the existing
positive legend assertions. Add the A2UiLegend import from a2ui_panels.dart and
use it to verify the showLegend condition.

In `@workout-logger/test/genui/components/scatter_plot_test.dart`:
- Around line 89-108: Extend the hostile-input tests around parse to include
points whose x or y coordinates are "NaN" and "Infinity" strings. Assert that
the resulting bounds contain only finite values, ensuring non-finite parsed
doubles are rejected before entering the bounds calculation.

In `@workout-logger/test/genui/components/stat_card_test.dart`:
- Around line 53-56: Add a test case in the “appends a unit that is not already
present” test covering a unit that appears only as a substring, such as value
“10 reps” with unit “s”, and assert the intended formatted result so the
behavior described in stat card parsing is pinned.
- Around line 11-23: Update the test helper pump to pass its props argument into
A2UiProps instead of constructing empty properties, so StatCardSpec.render
receives the caller’s values. Then replace the duplicated pumpWidget setup in
the title/value/subtitle test with a direct pump call using the intended
StatCard properties.

In `@workout-logger/test/new_features_test.dart`:
- Around line 192-231: Extend the CoachToolService test group with cases
covering get_health_metrics, analyze_health_workout_correlation, and
get_muscle_group_volume. Add assertions for the _hh == null error path,
insufficient paired-data behavior, and muscle-group lookup using a display name,
ensuring each test exercises the relevant days window, synthetic-data fallback,
and muscle-id matching defects without changing the existing sleeping-HR test.
- Around line 99-104: Update the assisted pullups test around calculateVolume to
use different values for WorkoutSet.weight and WorkoutSet.assistWeight, while
preserving the expected volume based on bodyweight minus assistWeight plus any
extra weight, multiplied by reps. Choose values that would produce a different
result if weight were incorrectly used in place of assistWeight.

In `@workout-logger/test/settings_provider_test.dart`:
- Line 23: Update SettingsProvider’s _getFallbackModel to use the supported
legacy model gemini-3.5-flash instead of gemini-3.6-flash, while preserving the
existing configuration and AI service behavior.

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In
`@docs/superpowers/specs/2026-08-08-sqlite-migration-and-coach-sql-tool-design.md`:
- Around line 189-191: Extend the migration flow in AppInitializer to validate
SQLite data before setting storage_migrated_v1: read back the migrated records
through exportAllData() and compare normalized data, identifiers, counts, nested
sets, JSON blobs, and settings against the Hive source. Set the Hive migration
flag only when validation succeeds; otherwise leave it unset and preserve the
existing migration retry behavior.
- Around line 128-133: Update the personal_records schema so records are keyed
by both exercise_id and handle, and define how existing legacy records without
handles are mapped during migration. Ensure migration preserves separate records
for multiple handles, and add coverage validating that no records collide,
overwrite, or get discarded.
- Around line 186-192: The migration flow in AppInitializer must be atomic and
safely retryable: execute all Hive reads and SQLite writes within one SQLite
transaction, or stage them in a fresh database and replace the active database
only after complete success. Ensure any failure discards all partial writes so
the next launch can rerun from current Hive data without stale rows or
conflicts, and add failure-injection coverage for retries after each entity
boundary.
- Around line 193-194: Update the Hive backup policy in the migration design to
define explicit retention and deletion behavior. Choose either removal after
verified migration with a documented recovery window, or encryption with
inclusion in deletion, reset, and export flows; replace the indefinite-retention
statement and apply the same policy to the related backup lifecycle section.
- Around line 172-180: Add an explicit schema version to
SqliteStorageService.init() by opening the database with version 1 and an
onUpgrade migration callback. Define the migration structure for future schema
changes, and include supported downgrade handling or an explicit data-preserving
down-revision path if sqflite provides downgrade support.
- Around line 204-206: Update the model query execution design to validate
row_limit within 1..500 before applying the outer LIMIT, and add tight bounds
for query length/shape and SQLite execution time to limit unbounded work from
scans, sorts, joins, or recursive CTEs. Preserve the existing error contract by
returning {'error': message} for cap or execution violations.
- Around line 68-72: Update the schema definitions for
exercise_muscle_activations.muscle_group_id, routine_exercises.exercise_id,
sessions.routine_id, exercise_logs.exercise_id, and targets.exercise_id to
include references to their corresponding tables. Also update onConfigure for
every writable connection to execute PRAGMA foreign_keys = ON.
- Around line 203-205: The connection design must not rely on
openReadOnlyDatabase as an independent safety boundary when it can reuse a
writable same-path instance. Update the read-only connection design to open with
singleInstance: false and require an Android integration test confirming writes
fail, or instead use an immutable snapshot/read-only-only handle enforced by
native SQLite.

In `@workout-logger/lib/services/ai/coach_tool_service.dart`:
- Around line 324-330: Remove readiness_score from the metric schema and all
related metric-handling paths in the coach tool service, including the synthetic
calculation near the readiness metric construction. Keep only measured health
metrics such as sleep_hours and deep_sleep_min, and update descriptions or
validation so readiness_score is no longer accepted or presented as independent
health data.
- Around line 1343-1347: Update _limitArg so the clamped value is explicitly
converted to int before returning it, while preserving the existing fallback and
1-to-max bounds.

---

Outside diff comments:
In `@workout-logger/lib/screens/widgets/exercise_input_section.dart`:
- Around line 129-177: Update the `_DropsetSection` call and implementation to
receive the assisted-bodyweight state from
`isAssistedBodyweightExercise(exerciseId)`. Use that state when rendering each
drop weight field so assisted exercises display “Assist (<unit>)” instead of
only the unit, and add help text explaining the assist value’s effective-load
calculation.

In `@workout-logger/lib/services/ai/coach_tool_service.dart`:
- Around line 560-604: The correlation flow around the sleepBars call must honor
the requested days window instead of always using HealthGranularity.week. Bound
days with _limitArg using the supported health-window maximum, then select the
appropriate sleep-bar granularity and trim/filter returned sleep data
consistently with _getHealthMetrics before pairing it with workout data.
- Around line 405-524: Update _getSleepingHrAnalytics to match the tool’s
declared output contract by calculating and returning median, P75, and the
complete declared aggregate percentile statistics, including corresponding daily
or series values where required. Reuse the collected sleeping-HR samples and
preserve existing fields, or revise the tool declaration to remove any
statistics that cannot be computed; do not leave declared metrics absent from
the response.
- Around line 574-599: Update the session aggregation loop in the
dayData-building logic to combine values for sessions sharing the same date
instead of overwriting m['y']. Accumulate workout_volume and session_duration
into the existing date value, while exercise_max_weight must retain the maximum
of the existing value and the current session’s max weight. Preserve the
existing exercise filtering and omit zero/unresolved maximum-weight results.

In `@workout-logger/lib/services/ml_service.dart`:
- Around line 375-407: Update the refSets initialization in the surrounding
method so that when lastSession is empty, it falls back to the first non-empty
session in pastSessions before the existing deload detection and empty-result
return. Preserve lastSession as the preferred source when available, and add a
regression test covering an empty lastSession with a valid historical session.

In `@workout-logger/lib/services/workout_provider.dart`:
- Around line 526-535: Update removeLastSet to preserve the current ExerciseLog
handle when rebuilding the log, matching the handle retention already
implemented in addSet. Pass currentLog.handle into the replacement ExerciseLog
so completed logs with remaining sets remain associated with handle-scoped
history.