# Tests and validation `test_sync_drive_transport` proves Drive artifact-before-manifest and manifest-before-reference publication, rejects unsafe/AI namespaces, and keeps a partial generation invisible. Orchestrator coverage proves that Drive mirror failure cannot undo a completed primary transport. `web_sessions` additionally proves duplicate-delivery rejection and stable-identity active-list removal after the reserved execution session appears in completed history. Android's normal unit suite builds the shared generation coordinator, foreground service and WorkManager transport rather than a test-only protocol copy. `web_sessions` exercises production preparation creation, durable request replay, immutable successor revisions, stale-revision conflict, duplicate exercise occurrences, list search, persisted detail serialization, revision- guarded withdrawal, conflicting idempotency-key reuse, active-list exclusion, and preservation of proposal/revision/delivery evidence. The HTTP test covers the real protected DELETE route and the versioned local preference GET/PUT, including CSRF, ETag, persistence and invalid-file fallback. The Web Vitest suite additionally covers cross-type global ordering beyond 32 rows, timestamp offsets, stable ties, invalid/absent dates, contextual states, confirmation cancel, the DELETE request and preference conflict behavior. Date tests run under `Pacific/Honolulu` and `Pacific/Kiritimati`. `SessionPreparationExchangeTest` covers Android import replay, occupied singleton refusal, explicit start, reserved execution identity, schema-v22 to v23 migration, permanent withdrawal replay, pending cancellation, started execution preservation, multi-delivery targeting, witness preservation and absence of performed facts. `test_session_preparation_export.py` proves that a withdrawn preparation is not redelivered, its exact evidence is exported in V2, and an acknowledged withdrawal is not republished. Generation tests bind the withdrawal to one immutable generation and acknowledge it only after the exact durable ACK. `SyncGenerationServiceTest` and `SyncDeploymentConversationTest` exercise the real optional artifact through generation validation and 24 bidirectional conversations. It also asserts that every foreground conversation emits exactly one fresh daemon request signal; the coordinator's production resume predicate preserves an interrupted generation identity until the counterpart generation has a durable terminal consumption row. `SyncDeploymentConversationTest` executes both production generation services and reopens Android storage. `test_web_sync_browser.py` adds a real Firefox click through the C adapter, observes the desktop commit at an Android test barrier, then releases the real ACK. This proves the directory transport double, not physical MTP. Orchestrator tests cover bounded streams and descendant cleanup; deployment-tool tests cover consistent SQLite backup and candidate inventory. The incident regressions also prove that a phase-owned MTP outbox excludes retained staging and unrelated legacy files, adapter expiry produces the stable `transport_timeout` code, and operational failure classes remain distinct. The C MTP regression keeps a malformed retained generation beside the current reference and proves that pull downloads only the referenced generation while still rejecting a truncated object inside that current generation. `web_session_deletions` exercises proposal withdrawal replay, preservation of a derived preparation, active-draft conflict, causal draft/session deletion and completed-session finalization. It also builds a Program-origin completed session, deletes it through the production command, verifies one causal operation/tombstone and terminal execution provenance, then replays the exact request. The HTTP suite covers real history detail, stale and successful session DELETE requests, invalid identity, post-delete deep-link absence and an injected SQLite failure with complete rollback. Vitest covers the focus-managed confirmation, Cancel and Escape zero-mutation paths, the real DELETE call, server-confirmed refresh, useful typed errors and focus restoration. AI companion tests prove that V2 exports a withdrawal without the proposal and that Android's tombstone rejects later V1 replay. `web_programs` exercises validation preview without writes, transactional import, exact replay, responsive detail serialization, archive replay, revision-guarded terminal deletion, response replay, preparation creation, usage/provenance and absence of delivery. Mixed-profile preparation coverage combines continuous duration, externally loaded reps with an integer JSON weight, and unloaded reps; it verifies preserved order and targets, generated entry identities, readable detail, request replay, and zero execution or delivery creation. A separate compatibility fixture retains an affected source Program's impossible zero weight while deriving an editable preparation with an unspecified target. Schema migration coverage checks all v26 tables and columns. `schema_v26_migration` additionally migrates a deterministically populated v25 fixture containing 24 sessions and 48 entries, with preparation provenance and history, and verifies exact typed projections after migration. Programs Vitest covers preview before the only importing mutation, explicit preparation creation/navigation, discrete trash confirmation, focus restoration, and Escape/Cancel zero-mutation behavior. The full HTTP, pagination, strict-parser and Core suites retain malformed, duplicate-key, bounds, catalog/profile/equipment, conflict and rollback coverage. Android Programs projection tests cover schema-v23 to v24 migration, full nondeleted `trainlog-programs` V1 snapshot import, durable deletion tombstones, read-only Sessions list/detail data, exact replay, capability gating, and correlated ACK behavior. The integrated `SyncGenerationServiceTest` invokes the real production chain from a live Program through capture, publication, Android consumption, deletion/tombstone publication, real ACK, desktop acceptance and reopen; an exact old generation is idempotently recognized and ACKed while the durable tombstone prevents Program resurrection, and a second exchange creates no duplication. Python Programs/synchronization suites cover the staged companion's full snapshot, unacknowledged tombstone publication, and ACK-driven acknowledgment without changing mobile V3, catalog, or preparation artifacts. Program execution regressions additionally cover Android schema v24 to v25, start/replay/singleton protection, target-only draft creation, completion, restart persistence and companion serialization. The integrated generation test uses the production desktop and Android producers/consumers to import a Program, execute one session, import completed history and its V1 provenance, acknowledge the correlated generation, expose Web `completed`, then continue through Program deletion, restart and old-generation replay without duplication or resurrection. The desktop populated migration test now continues through schema v27 and proves the new execution ledger starts empty. `schema_v28_migration` migrates a populated v27 fixture, preserves Program, session and execution identities byte-for-byte, admits the terminal deleted state, reports `integrity_check=ok` and an empty foreign-key check, and proves that migration invents no causal operation or deletion. The mobile V3 importer regression proves deterministic live causal revision seeding and exact replay. The Program execution exchange regression proves an old retained completed fact cannot override a session tombstone or duplicate terminal provenance. `test_web_sessions_browser.py` runs headless Firefox against the embedded production assets and C HTTP server with private temporary XDG roots. It creates only a synthetic ready preparation and delivery, verifies French then persisted ISO presentation, captures desktop and an exact BiDi-emulated 390x844 viewport, performs the real confirmation click, checks the schema-v24 withdrawal ledger after reload, and validates the resulting V2 export evidence. Android repository tests exercise the actual V2 consumer rather than a permissive protocol mock. `web_prepared_items` exercises the production SQLite projection and JSON serializer with a proposal and an execution draft, then proves finalized and causally deleted execution drafts are absent. The HTTP test covers the read-only route and method rejection. Vitest distinguishes proposal, active draft, loading, empty and failed-refresh states; it retains an already rendered durable value across a reread failure. Sync-control tests cover a single abortable polling scheduler, per-run revisions, rejection of an older run's late terminal response, stable error/action rendering, and cleanup on unmount. The `ai_history_export` test covers an empty database, a simple session, distinct multi-occurrences of one exercise, ordered sets, exact occurrence feedback, session-owned follow-up feedback, JSON escaping/UTF-8, MAX history, body measurements, deterministic content, and byte-for-byte database immutability. ## 1. Principle A Trainlog feature is not complete without relevant validation. Frozen formats, persistence migrations, synchronization semantics, and user-data mutations require executable coverage where practical. `WEB_CLI_HTTP_INFRASTRUCTURE_V1` adds pure CLI tests for the default TUI mode, both Web switches, help, version, unknown and repeated options, positional arguments, Web-only `--port`, missing/non-numeric/negative/zero/out-of-range ports, 65535, and option ordering. Process-level tests execute `--help` and `--version` without initializing the database, TUI or server. The isolated HTTP test opens one in-memory database, starts the production single-threaded adapter on an explicitly reserved loopback port, and verifies health JSON/version, content type, `nosniff`, CSP, absence of CORS, 404, 405 with `Allow: GET`, invalid Host, oversized body, oversized-header rejection, SIGTERM shutdown and explicit failure on an already occupied port. It never uses the Internet. A separate real-launch smoke uses an isolated XDG directory, performs the health request and stops the production binary with SIGINT. `WEB_FRONTEND_SHELL_V1` adds TypeScript typechecking and 10 Vitest/jsdom component tests for the persistent shell, five client destinations, active navigation, browser history, unavailable factual fields, successful/failed/ invalid health responses and accessible landmarks. Production Vite output is then generated and embedded for the C HTTP test, which additionally verifies HTML root and SPA routes, JS/CSS MIME, HTML and immutable cache policies, SHA-256 ETag, strict CSP, HEAD, unknown assets, traversal rejection and the invariant that unknown `/api/` paths never receive `index.html`. The generator contract test proves byte-for-byte deterministic output, stable path ordering and the 128-asset bound. Runtime smoke coverage starts the linked binary with Node/npm absent from `PATH`, removes any dependency on `web/dist`, and checks root, health and SIGINT. A separate `-Dweb=disabled` build verifies the explicit diagnostic instead of pretending that a frontend exists. Frontend preparation and direct validation are: ```bash cd web npm ci npm run typecheck npm test npm run build ``` Release-oriented Meson builds must use `-Dweb=enabled`; Meson never provisions or downloads npm dependencies. The history-deletion regression suite also exercises a real Program-origin completed session through the Core and HTTP command paths, including stale revision, cancellation, injected transaction failure, exact request replay and terminal execution provenance. Full-generation tests apply causal deletions before older history and equipment-association companions, prove that standalone imports remain strict, recover recent consumed and rejected ACKs without letting an older bounded window starve active producer capacity, and replay an already consumed generation without mutation. `WEB_DASHBOARD_DATA_CONTRACT_V1` adds a standalone bounded Core test for empty history, explicit unavailable domains, two observable sessions on one local day, an unfinished session, continuous activity, MAX-only sessions, primary zone deduplication, catalogue-without-history exclusion, 90/30-day windows, recent-MAX truncation/`partial`, Footer facts and JSON escaping/overflow. The HTTP lifecycle test additionally proves `GET /api/v1/dashboard`, JSON content type/version, unavailable reasons, corrupt-timestamp `invalid_data`, Core-error translation to bounded HTTP 500 and `POST` rejection without weakening Host, CSP, CORS or unknown-route behavior. Vitest validates strict parsing, real Footer population, empty fallback, malformed snapshots and absence of invented tile metrics; the seven tile bodies remain placeholders. `WEB_DASHBOARD_GRID_V1` adds Vitest coverage for the seven stable identities, default validity/non-overlap, every min/max constraint, independent rejection of unknown/duplicate/out-of-grid/non-integral layouts, lossless library adapter round trips, deterministic collision compaction, horizontal/vertical keyboard resize, responsive six/one-column projections, normal-mode locking, absent normal-mode handles, edit actions, cancel/reset/session-only save, visible keyboard focus and live announcements. Existing shell tests continue to prove the factual Footer, navigation and absence of invented data. `WEB_DASHBOARD_LAYOUT_V1` adds Core-independent filesystem tests for default, strict parsing, bounds, corruption, oversized/non-regular/symlink inputs, private permissions, atomic replacement, revision conflicts, preservation on invalid writes and reset. A canonical JSON parity test prevents C/TypeScript constraint drift. HTTP tests cover GET/PUT/DELETE, ETag/If-Match, two stale clients, CSRF/Origin rejection and `trainlog.perf` Origin acceptance. Vitest covers default/persisted/invalid sources, ETag/token parsing, canonical-only PUT, conflict translation, loading, local cancel/reset and durable Save. `WEB_DASHBOARD_TILES_V1` adds strict TypeScript parsing for every consumed field and Vitest fixtures confined to test sources. Coverage proves the 90-day activity map and direct totals, all compact/medium/large disclosure levels, legacy 0 kg visibility, `improved` use, unfinished-session duration absence, 1/3/8 MAX bounds, factual muscle bars, unavailable cardio/next-session states, global loading/error/invalid signals, accessible textual equivalents and live density changes during keyboard resize. `WEB_DASHBOARD_VISUALIZATIONS_V1` adds Vitest coverage for lazy chart presence by tile size, factual progression tooltips, one/multiple/legacy-zero points, accessible series descriptions, the complete canonical BODY ZONES SVG mapping, unmapped aggregate preservation, deterministic session-only intensity levels, front/back views, keyboard-focusable regions and factual muscle tooltips. The existing Activity, MAX, session, unavailable, Footer and Grid/Layout tests remain unchanged and passing. `WEB_DASHBOARD_V1_FINAL_REVIEW` adds explicit value-axis coverage for an empty defensive domain, one point, identical points, low amplitude, normal amplitude and a legacy zero. Every observation must remain within the computed bounds and a positive constant series must have visible space on both sides. Tile-density, BODY ZONES text preservation, edit handles, keyboard editing and responsive non-persistence remain covered by the existing Grid and tile suites. ## 2. Frozen Trainlog JSON v1 Run: ```bash python tools/validate_json.py ``` It validates: - `examples/session-v1.json`; - all positive fixtures; - all negative fixtures. A negative fixture passes only when Trainlog rejects it. Android field-workflow regression coverage additionally permutes distinct and duplicate active occurrences, checks stable IDs/equipment/targets/actuals and raw form state, recreates the repository, injects a transactional write failure, and finalizes the reordered draft. Completed-session tests edit reps, loads, set and continuous durations, distance, and speed; preserve feedback/follow-up roots; exercise atomic rollback; and inspect the corrected V3 export identity. Completed-order coverage also exercises duplicate exercises, SETS, CONTINUOUS, MAX, planning, equipment, feedback revisions, J+1 follow-up, Cancel/no-write, successful position persistence, injected Save rollback, and desktop V3 reconciliation without a duplicate canonical session. Field-input coverage maps one accumulated drag directly from index 5 to 1 and between both list extremes, restores the origin permutation on cancellation, and checks decimal editing/final parsing for both separators. It covers `0.3`, `0,3`, quarter-unit values, whole values, incomplete separator states, malformed text, non-finite spellings, negative final distance, stored numeric equality, and locale-independent V3 JSON publication. Semantic validation includes: ```text stable-ID uniqueness normalized exercise-name uniqueness timestamp offsets end > start catalog/reference equality tracking-mode consistency load-mode/weight consistency unknown-field rejection non-blank notes zero actual repetitions allowed ``` ## 3. Import reconciliation contract Run: ```bash python tools/validate_import_contract.py ``` Coverage includes: ```text same ID + same name/profile -> reuse same ID + renamed display text -> controlled reuse/warning same ID + incompatible mode -> reject different ID + equivalent normalized name -> reject new unique identity -> create ``` That command validates the frozen Trainlog JSON V1 importer contract. Active mobile V3 synchronization has a separate, stricter safe-reconciliation policy covered below; it does not modify the frozen V1 expectation. ## 4. Desktop Meson suite Current registered set (Meson executes tests in parallel, so displayed order is not contractual): ```text database catalog equipment_catalog body_zones training_knowledge session_generation session_generation_policy_validation training_context custom_equipment session_detail duration body_metrics bodyviz exercise_performance session_type_schema session_edit body_observation_edit usb mtp exercise_profile_schema continuous_session continuous_detail reps variable_sets schema_v5_migration schema_v9_migration schema_v7_migration timestamp_validation mobile_import_variable_sets mobile_import_multi_occurrence session_exchange_v3 equipment_associations_exchange equipment_definitions_exchange exercise_reconciliation body_zone_sync body_zone_catalog_validation sync_direction sync_history sync_screen_action measured_max max_results max_sync body_analytics terminal_input_event_type_policy tui_workflows app_shell presentation_i18n ``` The desktop suite includes APP_SHELL_V1 production-transition coverage and the AI-draft exchange regressions. The configured suite size and latest result are recorded once in `current_state.md`; the following generator-specific checkpoint counts remain historical evidence: ```text 45/45 Meson tests PASS ``` `SESSION_GENERATOR_V1` validation covered policy shape and shared fixtures, full-history exposure/recency boundaries, selection and observed-load anchors, Android v10 -> v11 planning migration, V3 round trips and malformed-V3 priority, and ordinary draft/editor acceptance. The checkpoint also passed 4/4 selected ASan/UBSan tests and Android 73 tests with zero failures/errors; one known historical real-v9 fixture skipped while the structural v10 migration test executed and passed. No hardware MTP or real app-upgrade/install validation is claimed for this checkpoint. The completed post-repair matrix passed 45/45 Meson tests, named ASan/UBSan 4/4, and Android 75 tests with zero failures/errors and one unavailable external Android-v9 fixture skip. The structural Android v10 planning migration executed and passed. Validators, strict C17 headers, deterministic policy/ fixture/knowledge regeneration, and APK asset byte comparisons passed. The earlier sanitizer invocation that selected no tests is not used as evidence. The desktop executable is additionally smoke-checked in isolated tmux PTYs at 100x30, the exact 72x20 minimum, and the 60x15 fallback; a resize down/up must recover before a clean keyboard quit. Notcurses is verified as the executable's direct terminal dependency with `readelf -d`. `presentation_i18n` covers the deterministic French default, French/English catalogue lookup, immediate in-process language switching, selected-language decimal/date formatting even under a contrary host numeric locale, unchanged layout at the 72x20 minimum, and local XDG configuration persistence/failure semantics. It also proves that persisted catalogue names are not translation keys, the source-derived `TRANSLATABLE_UI=0` boundary, and that synchronization presentation uses local typed status/counters rather than raw opaque core/receipt/history summaries. Android Robolectric coverage verifies both resource locales, Settings language ownership and dedicated SharedPreferences persistence, French fallback for an absent/invalid preference, and localized presentation of known stable BODY ZONE IDs without changing unknown catalogue labels or repository values. Resource parity and typed synchronization-receipt presentation are also covered; raw protocol/operational summaries remain opaque. Final i18n validation passed 60/60 normal desktop tests, 60/60 ASan/UBSan desktop tests, and 202 Android tests (198 passed, 4 skipped, 0 failed), with both JSON validators passing. Automated coverage does not replace a real device/manual visual language-switch smoke; that remains explicit manual validation. `WEB_DASHBOARD_CHARACTERIZATION_V1` now exercises the production Core service directly through a real in-memory SQLite database, without including `tui.c` or linking Notcurses. Its explicit fixed query instant covers an empty training history, exact rolling-window inclusion immediately before/at/after the 30-day boundary, valid actual history without `ended_at`, and plan-only exclusion. It covers repeated occurrences, heterogeneous REPS sets, SETS/DURATION, CONTINUOUS/DURATION without fake sets, assistance, unweighted actuals, ordinary heavy sets versus explicit MAX, equal-instant tie-breaks (including equal same-session performance points), the legacy unweighted `0.0 kg` prior baseline, primary-only catalogue zones, secondary-zone exclusion, and seeded catalogue behavior. Exact and exceeded bounds are asserted for 128 observable sessions and 4096 Dashboard facts, including the independently truncated totals and `partial` flag. Existing regressions continue to cover malformed legacy timestamps, body-metric selection, responsive rendering, all-history buckets, strictly earlier exact-dose improvement, and explicit-MAX separation. The standalone `dashboard_header` target also proves that `trainlog/dashboard.h` compiles independently as C17. TUI workflow regressions separately verify service adoption and snapshot presentation. Notable regression coverage: - transactional persisted-session replacement; - exercise removal from a session; - body-observation stable-identity editing; - profile-aware exercise constraints; - continuous activity without fake sets; - table-only desktop SETS workflow: planning does not create actual rows, explicit add requires an actual metric, and normal zero-row completion is rejected; Android legacy compact-draft decoding remains separately covered; - direct v4 -> current database migration and v7 -> v8 custom-equipment migration; - bounded v8 -> v9 explicit-max migration, including ambiguous-attempt preservation; - lossless v9 -> v10 `performed_sets` rebuild: historic NULL and positive weights/IDs/owners/positions/metrics survive, explicit zero is accepted, and injected failure rolls back with integrity, foreign-key and enforcement checks; - additive desktop v10 -> v11 and Android v9 -> v10 body-zone migrations: exact stable row identities and history remain unchanged while only proven manifest mappings are seeded; the Android fixture also preserves an active draft, per-set loads, explicit MAX, a body observation and a custom equipment definition; - canonical body-zone validation rejects duplicate IDs/order, missing parents, cycles, group/leaf disagreement, a taxonomy other than the exact V1 manifest, invalid `full_body`, boolean versions, malformed stable IDs, and direct group assignment; - one primary/multiple secondary persistence, duplicate/unknown/group/orphan- secondary rejection, primary/secondary exclusivity, edit, unclassified and reopen; - exact and parent-descendant zone filters, lower-body leaf coverage, primary-only participation, normalized-prefix composition and custom rows; - body-zone companion export/import, empty mapping, identical replay, one-sided update, simultaneous conflict rollback and source-V2-proven exercise-ID reconciliation without name-only inference; custom-exercise publication establishes both peer baselines before a reverse one-sided edit; malformed `ex_` identities and timestamps without offsets are rejected; - heterogeneous mobile-set import; - per-set load persistence and correction: ordered rows retain mixed actual repetitions, nullable loads, positions and assistance semantics through desktop replacement and detail retrieval; - V2 mobile round-trip and idempotent replay preserve each ordered set's own nullable `weight_kg`, without collapsing it to an occurrence target; - V2 explicit-max Android -> desktop -> Android replay and resumed same-session update; - Notcurses input lifecycle translation: PRESS/REPEAT are actionable while a RELEASE event is consumed without creating a second navigation action. - targetless mobile SETS persistence; - mobile-import idempotence. - V2 mobile import with repeated exercise occurrences and stable `entry_id`; - companion equipment import after session import, including idempotent reimport and independent associations for repeated occurrences; - rejection of the historical equipment-import invocation without its required `--database` target, followed by the corrected complete V2 export chain. - stable-ID mobile-to-desktop rename reconciliation without duplicate catalog rows or historical-reference replacement. - supplied catalogue search/detail, local custom-equipment creation and selection, resolution of supplied/local occurrence links, and explicit visibility of an unknown historic equipment reference; - strict custom-equipment definitions V1 import/export: reserved supplied IDs, additive omission, equal reimport, and divergent same-ID conflict; - three-mode shared-engine plans and their directional receive/publish bounds; - Sync-page action dispatch: direct `a`/`p`/`b` confirmation, one `Enter` run, `Esc` cancellation without a run, refresh-only `r`, inert retired `s`, and equivalent full/compact footer direction labels; - definition-first ordering before V2 artifacts, so associations can resolve custom IDs without changing V2 shapes. - exercise reconciliation with identical profiles, compatible subset/superset masks, explicit incomparable-profile rejection, current Marche identities, preservation of two historic occurrences and their `entry_id`/continuous values, ordered custom-equipment definitions/catalog/mobile/association replay, complete definitions/mobile/associations/body import, outbound publication, and stable second replay; - a production PC-exporter to Android-importer regression over definitions, catalog, mobile V3, and association V2 artifacts: the fixture includes Marche, Leg press, two Marche occurrences, per-set loads, body data, a durable draft, and custom equipment; after the first import, both the second and third imports report zero additions and an exact snapshot of every Android business table remains unchanged; - synchronization summary reporting where a reconciliation-only result keeps `+0 exercice(s)` and only an actual inserted row produces `+1`; - all fourteen body metrics through the same complete reconciliation import; - desktop-created occurrence identities use `sxe_`, not the `sy_` synchronization-run prefix. ## 5. Build ### Isolated synchronization validation environment `TRAINLOG_SYNC_TEST_ENV_V1=PASS/FROZEN` provides one repository-owned entry point for the desktop and Android validation needed by synchronization work: ```bash python3 tools/validate_sync_isolated.py --suite preflight python3 tools/validate_sync_isolated.py --suite smoke python3 tools/validate_sync_isolated.py --suite full --jdk /usr/lib/jvm/java-17-openjdk ``` `TRAINLOG_SYNC_DATA_LIFECYCLE_V1` adds schema v19/v18 migration assertions and Android lifecycle coverage for confirmed-only serialization, persistent pending state, occupied-singleton refusal, explicit activation, identity- preserving idempotent finalization and stale replay. The closeout adds `v4HistoryRoundTripsAcrossRealAndroidAndDesktopImplementations` and `executionDraftRoundTripsAcrossRealAndroidAndDesktopImplementations`. Their bridge provisions only synthetic schema and invokes the real Android and desktop staged entry points; it does not reconstruct wire data. The final isolated run passed 75/75 normal native tests, 75/75 ASan/UBSan native tests and 207 Android tests (203 passed, four unchanged historical-fixture skips), plus both Python validators and `assembleDebug`. These are artifact/repository proofs; they do not claim transport activation or global synchronization atomicity. The 75-versus-73 inventory contains no removed, renamed, merged, or disabled native tests. With `web=auto` and no `web/node_modules`, Meson selects 73 tests and omits exactly `web_frontend_typecheck` and `web_frontend`. After `npm ci` from the locked package file and Meson reconfiguration, both the normal and ASan/UBSan selections contain and pass all 75. Clang 22.1.8 is the compiler used for those green builds. GCC 16.2.1 fails both baseline `77c1517` and the lifecycle branch under the unchanged strict `-Werror` policy on `-Wmisleading-indentation` in unchanged `json_writer.c`, `web_dashboard_json.c`, and `web_dashboard.c`; this is a pre-existing compiler-version limitation, not lifecycle evidence and not a relaxed gate. Use `--jdk /absolute/jdk17/home` when auto-detection is unsuitable and `--run-parent /absolute/private/parent` to select another parent outside the problematic system `/tmp`. `--keep-run-dir` retains a successful run for diagnosis; failures and interruptions are always retained. Without that flag, a successful run removes only its own validated `run-*` directory. Each invocation creates a mode-0700 unique root under `${XDG_CACHE_HOME:-$HOME/.cache}/trainlog/test-runs/`, then gives its children private HOME, XDG data/config/cache/runtime, native/JVM tmp, exchange, database and report directories. Test-sensitive files and logs are mode 0600. The already-provisioned `${GRADLE_USER_HOME:-$original_HOME/.gradle}` dependency cache is deliberately shared; no Trainlog application data, rclone credential, keystore or production XDG path is shared. This is path/environment isolation, not a system sandbox. The harness verifies Java 17 with `java -version`, then verifies both Gradle's launcher and daemon JVM. `JAVA_TOOL_OPTIONS=-Djava.io.tmpdir=/tmp` reaches the forked Robolectric JVM; `IsolatedTestEnvironmentTest` checks the effective property and every private path from inside that JVM. Android test tasks use `--rerun-tasks`, so an up-to-date or cached task cannot masquerade as a new execution. No default mode invokes rclone, adb, libmtp synchronization, `trainlog-sync-once`, `connectedDebugAndroidTest`, a user service or a production lock/database. The modes are: - `preflight`: tools, writable parent, JDK 17, Gradle launcher/daemon and paths; - `smoke`: preflight, production-path sync-gap characterization and the Android JVM isolation test; - `full`: normal and ASan/UBSan desktop suites, explicit characterization, JSON/import validators, targeted draft lifecycle/export test, complete Android unit suite and `assembleDebug`. Logs are private and capped at 8 MiB per command. Exit codes distinguish `0` success, `2` blocked preflight, `3` missing dependency, `4` environment failure, `5` business-test/command failure and `130` interruption. Commands, exit codes, JDK, non-secret paths and Android test/pass/fail/skip counts are recorded in `reports/`; the full environment and credentials are never printed. The harness terminates only process groups that it started, waits before forced termination, never stops a pre-existing Gradle daemon or Trainlog service, and refuses cleanup outside its exact root or across a symlink. At the 2026-09-17 reference checkpoint, the Android total retained four skips because optional historical database fixtures were unavailable: one v9, one v12, one v13, and one v14 migration-fixture case. These are explicit fixture coverage gaps, not passing migration executions. The isolated harness does not run hardware MTP, real Drive, or Android instrumented tests; those remain separate validation classes. Stable-release validation additionally builds the signed release variant and checks the packaged product versions and binary linkage: The release machine supplies all four `TRAINLOG_RELEASE_*` environment variables, or a mode-0600 private file at `~/.config/trainlog/release-signing.properties` with this shape: ```text storeFile=/absolute/private/path/trainlog-release.jks storePassword= keyAlias= keyPassword= ``` `TRAINLOG_RELEASE_CREDENTIALS_FILE` may select another private properties file. Environment variables take precedence over file values. The keystore and credentials are the durable Android update identity: back them up through an approved encrypted external mechanism, never Git or release assets. Missing configuration makes `packageRelease` fail explicitly rather than producing an apparently publishable unsigned APK. ```bash meson compile -C build meson test -C build --print-errorlogs cd android JAVA_HOME=/usr/lib/jvm/java-17-openjdk ./gradlew test JAVA_HOME=/usr/lib/jvm/java-17-openjdk ./gradlew assembleRelease apksigner verify --verbose --print-certs \ app/build/outputs/apk/release/app-release.apk file trainlog-tui-linux-x86_64-v ldd trainlog-tui-linux-x86_64-v sha256sum trainlog-android-v.apk \ trainlog-tui-linux-x86_64-v ``` An unsigned release APK is build evidence only and must not be published as the canonical stable Android artifact. Release assets must come from the exact tagged source state, and the tag, Android `versionName`, Meson project version, TUI product version, asset names, and release title must agree. Private update rollout may set `TRAINLOG_ANDROID_VERSION_CODE` to a positive integer above the installed package while retaining the current development `versionName=0.1.3`. The version-code override is build-only and does not change any schema or exchange-format version. Private daily candidates must not downgrade the logical product version to the earlier 0.1.2 stable line. For `TRAINLOG_UI_ANDROID_WEB_UNIFICATION_V1`, `lintDebug` was also run from a detached clean worktree at the exact `main` baseline and from the mission branch. Both runs report the same ten errors: one `NewApi` diagnostic for `windowLightNavigationBar`, one `LocalContextConfigurationRead`, and eight `StringFormatMatches` diagnostics (four bilingual call sites). Comparison by rule and complete diagnostic message reports zero added and zero removed errors. No lint rule or severity was changed. ```bash meson setup --reconfigure build meson compile -C build meson test -C build --print-errorlogs ``` Strict warning flags remain active. Do not weaken warnings to make a change pass. ## APP_SHELL_V1 production-transition coverage The active desktop suite has 58 tests. `app_shell` verifies layout thresholds, route/history bounds, overlays and focus restoration, bounded UTF-8 search and form input, stable-ID list selection, shared action ordering, and leave guards. `tui_workflows` covers the production single event-loop routes and controller handoffs: sessions/generator, exercises, equipment, statistics/body/MAX, settings, and confirmed synchronization. `custom_equipment` covers the bounded deterministic read-only equipment page reader, including pagination, invalid arguments, offsets and corrupt values. This coverage replaces former nested-screen-loop workflow claims. The Mensurations workflow regression separates the selected observation's profile from its metric history: singleton circumference bars, independent kg/cm rendering, missing-field omission, a real two-date high-resolution Unicode series, historical selection updates, availability-filtered metric navigation, and bounded 120x35, 100x30, 80x24 and 72x20 geometry without ASCII chart glyphs. The terminal-free `chart` regression exercises zero, one and two points; constant, increasing and decreasing series; irregular elapsed-time X mapping; Y headroom; clipping and compact rectangles; Braille 2x4 dot encoding; interpolated Braille output; and distinct stored-measurement markers. The terminal-free `statistics` regression covers empty, 7-day, 30-day and complete windows; actual-session and occurrence counts; sets, repetitions and duration; strict external-load volume; same-occurrence planned/actual totals; canonical alias reconciliation; primary and secondary persisted zones; immediate and session-global feedback scopes; and zero/one/multiple explicit MAX histories. `tui_workflows` additionally exercises the five-entry statistics hub, period and graph cycling, compact zone bars, canonical exercise detail, MAX facts, and bounded common-chart rendering without an interactive terminal. Statistics-refinement regressions distinguish identical observations plotted inside 7-day and 30-day domains, clamp non-negative scales to zero, render two weekly totals as independent bars, retain empty local-calendar weeks, compare unequal zone counts through one shared scale, and verify that the default MAX list excludes never-measured exercises until its explicit toggle is used. Period-bucketing regressions cover four consecutive 7-day comparison buckets, Monday/Sunday assignment, localized consecutive Gregorian months, retained empty middle periods, and boundary assignment without double counting. Ordinary February, leap-year February and exact first/last dates verify month length without fixed-day arithmetic. Offset-bearing fixtures verify persisted civil-date membership independently of the machine timezone. Current partial weeks/months, stable zero history, and duration-based adaptive full-history aggregation are also covered. The APP_SHELL PTY validation exercises six TUI sizes—72x20, 80x24, 100x25, 100x30, 120x31 and 120x35—plus help, search clear/close, F6/F7, compact focus, resize/overlay restoration, clean exit, and navigation with no temporary database write. Normal and ASan/UBSan Meson suites each passed 48/48; the current normal and sanitizer real-PTY runs each passed 100/100 checks. The sanitizer run used the upstream-prescribed Notcurses compatibility setting `ASAN_OPTIONS=use_sigaltstack=0:detect_leaks=1:halt_on_error=1` and reported no ASan/UBSan diagnostics. Android unit/assembly evidence records 84 tests, 0 failures, 0 errors and one external `TRAINLOG_ANDROID_V9_FIXTURE` skip. Required human checks remain: the six TUI sizes above; Android widths 320, 360, 393 and 412 dp; large system font; IME forms; durable-draft leave guards; and TalkBack. No emulator was available; the daily installed application was not installed over or exercised by instrumentation, so these are not marked visually passed. ## 6. Android build When Android code changes: ```bash cd android printf 'sdk.dir=%s\n' "$HOME/Android/Sdk" > local.properties JAVA_HOME=/usr/lib/jvm/java-17-openjdk \ ./gradlew testDebugUnitTest assembleDebug ``` Install to the connected device when hardware behavior changes: ```bash adb install -r app/build/outputs/apk/debug/app-debug.apk ``` Android repository host tests additionally cover exercise editing: - trimmed rename preserves `exercise_id` and recalculates `normalized_name`; - duplicate and invalid names are rejected; - completed history and active-draft references resolve the renamed catalog row; - a referenced profile change is explicitly rejected; - same-ID PC-catalog rename reconciles in place without a duplicate. - canonical body-zone taxonomy, create/edit/reopen, parent filters, search+filter, unclassified rows, companion replay/update/conflict and zone-safe exercise-identity merging. `BodyZoneHomeOverviewRepositoryTest` exercises the production `getBodyZoneHomeOverview(now)` query with a caller-supplied deterministic instant. It covers the settled chest/back/shoulders/thighs example, secondary-only exposure remaining weaker than primary, stable ranking, unsupported calves exclusion, parent-zone non-participation, retired alias A to canonical B counted once, active/resolved availability, honest empty history, inclusive exact 7/30-day instant boundaries, and target-only exclusion. `BodyZoneHomePresentationTest` covers concise evidence and last-primary age wording without recovery/fatigue/prescription language. `BodyZoneHomeSemanticsTest` renders the production section at 288 dp content width (the constrained 320 dp-screen case) with 1.6x font scale and at a practical 380 dp content width. It verifies every view-specific Face/Dos zone and textual state, Button actions, selected/not-selected transitions, leaf canvases with no anatomical text children, exact normalized-path containment, a real Canvas tap selecting core, and both existing-route callbacks. The Home implementation keeps side-by-side maps with separate region-aligned, at-least 48 dp semantic controls; 320/360/393/412 dp and TalkBack visual checks remain explicit device/emulator validation. Focused host command: ```bash cd android JAVA_HOME=/usr/lib/jvm/java-17-openjdk \ ./gradlew :app:testDebugUnitTest \ --tests com.labfytools.trainlog.data.BodyZoneHomeOverviewRepositoryTest \ --tests com.labfytools.trainlog.ui.BodyZoneHomePresentationTest \ --tests com.labfytools.trainlog.ui.BodyZoneHomeSemanticsTest ``` ## 7. Hardware MTP validation Hardware probes and real synchronization are separate from the normal automated suite because a test runner cannot assume an unlocked MTP phone. Available probe binaries include: ```text trainlog-usb-probe trainlog-mtp-probe trainlog-mtp-exchange-probe trainlog-mtp-roundtrip-probe trainlog-mtp-mobile-export-probe ``` Previously established physical baseline: ```text USB_MTP_DETECTION=PASS MTP_STORAGE_ACCESS=PASS MTP_WRITE=PASS MTP_LIST_FOLDER=PASS MTP_READ=PASS MTP_ROUNDTRIP=PASS ``` The automated `sync_direction` regression also covers selection of the newest Android V2 export, equipment-definitions V1 artifact, and equipment-association V2 artifact when scoped storage has retained collision-suffixed sibling files. It rejects non-versioned/ non-numeric names and uses deterministic ties; JSON validation remains part of the normal import path. `mobile_import_multi_occurrence` exercises the real definition importer before the V2 session importer, including idempotent reimport and rejection of a missing custom definition. ## 8. Bidirectional synchronization validation `sync_gap_characterization` uses the production mobile V3 exporter/importer and temporary SQLite databases to prove the current gaps without blessing them: V3 omits `ended_at`, three distinct note owners and the body-observation session link; the imported values are `NULL`; deleting a local session and replaying an old snapshot resurrects it; an invalid artifact rolls back while an earlier artifact remains independently committed. The existing Android test `TrainlogRepositoryDraftTest.finalizeIsAtomicAndDraftNeverExportsBeforeCompletion` proves that the active draft is absent before finalization and the completed session appears afterward. These tests do not implement the future contract. `TRAINLOG_SYNC_CHARACTERIZATION_V1=PASS/FROZEN` completes the missing current- behavior evidence without changing production code, schemas, or formats. The coverage matrix is assertion-based; “future” rows remain criteria rather than passing tests: | Current behavior | Test and production boundary | Significant assertions | Coverage | |---|---|---|---| | V3 omissions, new-row loss, replay resurrection, and per-artifact rollback | `tests/test_sync_gap_characterization.py` through the real desktop V3 importer/exporter | Wire fields are absent; newly imported columns are `NULL`; an identical replay preserves local-only values; a V3 correction rebuilds its occurrence and loses its local-only occurrence note; replay after deletion resurrects the session; malformed input rolls back only its artifact | Existing proof extended | | Active draft lifecycle | `TrainlogRepositoryDraftTest.finalizeIsAtomicAndDraftNeverExportsBeforeCompletion` through the real Android repository/exporter | No draft rows export; atomic finalization removes the draft and makes the completed session exportable | Existing proof reused | | Cross-implementation V3 round trip | `TrainlogRepositoryDraftTest.androidV3RoundTripUsesRealDesktopImporterExporterAndFreshAndroidDatabase`, with `tests/support/roundtrip_desktop_v3.py` invoking the real desktop tools | A fresh Android database reconstructs stable session/entry/exercise/equipment identities, repeated occurrences, occurrence and set order, heterogeneous sets, targets, continuous values, and explicit MAX; exact replay adds nothing | Added | | V3 correction and forbidden rebinding | Existing desktop `session_exchange_v3` and Android repository cases | Accepted same-identity correction is reported separately from insertion; conflicting `entry_id` rebinding is rejected | Existing proof reused | | Companion identity and revision chains | Existing aliases, equipment, BODY ZONES, causal profile-state, and feedback suites | Stable foreign references, immediate-feedback roots, immutable revisions, session follow-ups, and replay behavior remain exact | Existing proof reused | | Inter-artifact partial failure | Existing Android inbox ordering tests and `sync_gap_characterization` | Earlier valid catalogue/components remain committed when a later session/companion is invalid; intra-artifact invalid content rolls back | Existing proof reused | | Desktop delete-before-send publication | `sync_body_zone_wiring` through `trainlog_sync_run()` and the MTP double | The old remote catalogue is deleted, the injected send fails, the run returns an error, and no replacement exists | Added | | Selected-invalid V3 priority | `TrainlogRepositoryDraftTest.inboxPresentMalformedV3DoesNotFallBackToValidV2` through the Android inbox importer | A present malformed V3 fails and is not masked by a valid V2 | Existing proof reused | | Request, receipt, and processed marker | `sync_body_zone_wiring` through `trainlog_sync_run()` | Receipt and report carry the exact `request_id`; identical request replay is refused; receipt-send failure leaves the prior processed marker unchanged | Added | | Android post-receipt consumption | Desktop request/receipt proof plus Android typed receipt/inbox component tests | Desktop processing success and Android artifact-import success are separate component results; V1 has no durable peer-consumption acknowledgement | Component-scoped only | | Inter-process lock | `sync_body_zone_wiring` against the real `sync.lock` acquisition with private `XDG_DATA_HOME` | A pipe-coordinated child holds the lock; the second trigger returns conflict before any MTP probe; every child is joined | Added | | Enriched V4 completed history | `TrainlogRepositoryDraftTest.v4HistoryRoundTripsAcrossRealAndroidAndDesktopImplementations` through Android V4 repository methods and the real desktop V4 importer/exporter | Android → desktop → fresh Android and desktop → Android → fresh desktop preserve stable identities, order, heterogeneous sets, plans, equipment, causal null/empty/escaped UTF-8 notes and idempotent replay | Added cross-implementation proof | | Execution-draft V1 | `TrainlogRepositoryDraftTest.executionDraftRoundTripsAcrossRealAndroidAndDesktopImplementations` through Android repository methods and `execution_draft_exchange.py` | Both implementation directions preserve stable session/entry identity and confirmed facts; raw unfinished form text remains local; replay is unchanged and pre-finalization replay is stale after reopen/finalization | Added cross-implementation proof | | General tombstones, coherent generations, causal deletion, and consumption acknowledgement | Causal-deletion suites plus `sync_generation_exchange` and `SyncGenerationServiceTest` | Explicit staged production entry points pass cross-implementation, rollback, replay, bounds and restart assertions; automatic V3 remains unchanged | PASS/FROZEN staged evidence | The cross-implementation helpers only provision minimal production-compatible desktop schemas and launch production Python tools; they do not duplicate their business rules. Kotlin tests use real Android exporters and importers on distinct databases. Comparisons remove only volatile `generated_at`, canonicalize object-key and unordered top-level collection order, and retain occurrence/set array order as business data. Request/receipt evidence remains deliberately bounded: the native test proves desktop correlation, publication, replay handling, and marking order, while Android tests prove local receipt presentation and inbox failure behavior. A receipt still proves desktop processing only, not durable Android consumption. README, `architecture.md`, and `sync_exchange.md` were reviewed for this lot and need no correction; their current ownership and V1 limitations remain accurate. The 2026-09-17 closing run used all three documented harness modes with the explicit JDK 17 path. The final `full` mode passed 75/75 normal Meson tests, 75/75 ASan/UBSan Meson tests, both validators, the targeted draft lifecycle test, Android debug assembly, and 204 Android tests (200 passed, 4 historical fixture skips, 0 failed/error). No hardware, Drive, adb, instrumented test, or user service was exercised. Validated workflow: ```text Android -> Synchroniser maintenant -> unique sr_ request -> trainlog-syncd -> shared engine -> Android -> PC import -> PC -> Android catalog -> sy_ structured run -> matching receipt -> Android final status ``` Multiple distinct Android request IDs were processed successfully without reprocessing one request as a new one. The TUI also invokes the same engine manually and exposes structured run details. `post_sync_ai_drive` validates strict export-before-upload ordering, skips upload after export failure, and reports unavailable/failing rclone. The production sync wiring test exercises successful TUI and Android triggers through `trainlog_sync_run()`, uses an isolated fake rclone, and proves that failed sync runs do not invoke Drive publication. Post-sync failures are non-fatal to the already successful Trainlog sync by contract. `ai_session_draft` validates the strict desktop source parser and schema v18 transaction: exact/unknown-key rejection, `aid_` identity, alias resolution, target-only SETS compatibility, 64-entry/99-set/999-rep bounds, exact replay skip, changed-payload conflict, deterministic 256-at-a-time publication across 257 drafts, failed-publication retry, permanent import ledger, and full rollback on an injected child failure. Its fake-rclone seam also verifies that the exact fetched bytes are archive-copied only after import, same-name inbox replacement is preserved, stale temporary content cannot import, and archive/status failure converges through idempotent logical replay. Production orchestration tests retain `DRIVE_FAIL`, fetched-invalid `REJECTED`, and `ARCHIVE_FAIL` in a successful report/history. Android repository tests cover the 256-proposal companion limit, full validation before replay skip, strict optional-text trimming, replay idempotence, start/delete tombstones, target-only copy to the unchanged singleton active draft, and catalog/profile/equipment prerequisite ordering. Focused Android UI wiring tests also cover pending-only deterministic listing, started/deleted exclusion, the always-visible **Brouillons** count without losing manual/history actions, proposal title/target details, inert display with no automatic active draft, external-revision refresh after companion import, and the existing transactional start path. These are automated fake transport tests; no real Drive automated test exists. The final real Drive plus Android-triggered bidirectional smoke test remains manual and `TRAINLOG_AI_SESSION_DRAFT_V1=VALIDATION_PENDING`. A September 2026 connected-device inspection established that the previously installed 2026-09-13 APK still owned a schema-v16 database without `ai_session_drafts`; it could not import the companion. Installing the current APK and synchronizing once is therefore a prerequisite for the remaining real acceptance test, not evidence that the acceptance test has passed. ## 9. Sanitizers For meaningful C checkpoints: ```bash CC=clang meson setup build-asan \ -Db_sanitize=address,undefined \ -Db_lundef=false meson compile -C build-asan meson test -C build-asan --print-errorlogs ``` ## 10. Pre-push checklist ```bash meson compile -C build meson test -C build --print-errorlogs python tools/validate_json.py python tools/validate_import_contract.py git diff --check git status --short ``` When Android changed, add: ```bash cd android JAVA_HOME=/usr/lib/jvm/java-17-openjdk ./gradlew testDebugUnitTest assembleDebug ``` Documentation must describe the resulting state, not retain contradictory old `NEXT` checkpoints. ## 11. Measured-max regression The normal Meson suite contains: ```text measured_max ``` Coverage proves: - stronger ordinary training is ignored by measured-max classification; - only `max_test` sessions participate; - zero-repetition failed attempts are not promoted; - newest successful explicit test is the current measurement; - historical external-load record can remain older than current; - lower assistance is better; - no-load max tests compare actual reps/duration; - external working loads round to the configured increment; - working-load percentages reject assistance. - `%MAX` target calculators use the exact unrounded formula, enforce 1..100, exact equipment identity and external resistance, and reject assistance; - Android manual target-plan tests prove that calculation is read-only, actual set weights remain unchanged, existing target dose/rest survive direct-kg edits, and none removes the plan target; - the real TUI `/` path resolves through the registered action, focuses search, filters live, then follows clear-before-close Escape semantics; - action stable identifiers are unique, preventing footer/F7 duplication. - Statistics-dashboard rendering covers global summary/event facts, primary-zone and unclassified active-catalogue projection without merged-alias double counts, zero/one-point without-graph and multi-point Unicode detail-chart states; readable selected-period frequency with non-duplicating MAX markers; daily/multi-day period projection and independent catalogue bars with right-side counts; exact-dose per-set parity, equal-instant working/MAX non-events, malformed-timestamp availability, distinct wide versus stacked compact geometry, and Unicode blocks used only as chart geometry; and footer F6/F7 single-occurrence output at 120x35, 100x30, 80x24, and 72x20. Historical validation checkpoint (not the current suite size): ```text 39/39 Meson tests PASS ``` ## 12. Body analytics regression The Meson suite adds: ```text body_analytics ``` Coverage includes: - male circumference-formula branch; - female circumference-formula branch; - estimated fat and lean mass from real body weight; - waist/hip, shoulder/waist, and chest/waist ratios; - left/right asymmetry; - profile-independent analytics without a configured estimation profile; - missing required circumference handling; - invalid estimation-profile rejection. Historical validation checkpoint (the current desktop suite is 48/48): ```text 39/39 Meson tests PASS ``` ## 13. Android session draft v1 Android schema v4 introduced one durable active draft; the current additive chain reaches schema v17 without clearing completed history or the draft. The explicit v10 -> v11 migration adds optional planning metadata while preserving existing rows with `load_mode=none`, zero rest and NULL targets. The v11 -> v12 migration adds the durable flattened exercise-alias table. Its exact physical-v11 fixture compares every pre-existing table cell before and after migration, requires the new alias table to be empty, and checks foreign keys. The dedicated physical v16 -> v17 fixture preserves representative completed history and the active draft, adds empty AI proposal tables, and checks SQLite integrity and foreign keys. AI companion parser coverage accepts title/notes at their 120/2,000 Unicode-code-point limits (including non-BMP characters) and rejects 121/2,001 code points atomically. Host coverage includes exercise shapes and raw partial text, fresh repository restore, remove/discard, atomic finalization and repeated-finalize rejection, rollback, catalog reconciliation, missing-selection recovery, explicit DB-open failure, historical migration, equipment selection and occurrence identity. Schema-v9 host coverage additionally proves French raw max-text persistence, explicit max creation/edit/finalization without sets, distinct movement values on the same equipment, latest-per-exercise history, V2 replay, stable-ID resume and bounded conversion that leaves multiple legacy attempts untouched. Schema-v10 host coverage adds the canonical body-zone asset, stable-ID mapping, relation constraints, filters, transactional editing and companion conflict policy without changing explicit MAX or per-set tables. The current set-row-editor coverage additionally proves that row edits, deletion and append preserve neighbouring rows; French-comma loads, blank loads and explicit zero loads remain distinct; reopening preserves aligned raw row fields; and completed history keeps the ordered per-set values. The instrumentation source exercises row edit, deletion, append and Activity recreation, but this document does not claim that instrumentation was executed for the current documentation checkpoint. For the settled implementation, the validation inventory is: ```text Android JVM: ./gradlew testDebugUnitTest Android compilation: ./gradlew assembleDebug Android instrumentation: adb shell am instrument ... (execution is explicit; no device execution is asserted here) Desktop: meson compile -C build meson test -C build --print-errorlogs Frozen JSON/import: python tools/validate_json.py python tools/validate_import_contract.py V2 regression: mobile_import_variable_sets and the desktop/Android V2 round-trip/idempotent-replay coverage Sanitizers: clang ASan/UBSan Meson build and test invocation ``` Executed Body Zones V1 evidence is recorded after each closeout run: Android `testDebugUnitTest` 44/44 with the retained real v9 fixture enabled and `assembleDebug`, a historical 39/39 Meson checkpoint, valid and invalid JSON checks, import-contract checks, and the ASan/UBSan Meson suite. Device installation and installed Android-store migration remain explicit hardware steps and are never inferred from host tests. The 2026-09-09 desktop closeout additionally backed up the real v10 database, opened it through the production Notcurses binary, and verified v11 integrity, foreign keys, all historical row counts and bidirectional row equality against the backup. A real Kitty terminal validation exercised the zone list, parent filter, classified/unclassified details and edit preloading/cancellation. The Samsung SM-G990B then passed certificate matching, `adb install -r`, the real v9 -> v10 migration, SQLite integrity/FK checks, row equality for every pre-existing application table, the Android zone UI matrix and two live MTP round trips. The second run reported zero additions/reconciliations and left the Android application tables and semantic companion exercise states equal to the first run. A regression also covers a current `trainlog-sync-request-v1 (N).json` beside an older canonical request. The optional `RealAndroidV9BodyZonesMigrationTest` is enabled by setting `TRAINLOG_ANDROID_V9_FIXTURE` to a coherent copied v9 database. It makes two test-owned copies before opening the production repository, then compares all 16 pre-existing application tables in both directions; it never opens or modifies the supplied fixture through the migration helper. ```bash cd android JAVA_HOME=/usr/lib/jvm/java-17-openjdk ./gradlew test JAVA_HOME=/usr/lib/jvm/java-17-openjdk ./gradlew assembleDebug assembleDebugAndroidTest adb install -r app/build/outputs/apk/debug/app-debug.apk adb install -r app/build/outputs/apk/androidTest/debug/app-debug-androidTest.apk adb shell am instrument -w -e package com.labfytools.trainlog \ com.labfytools.trainlog.test/androidx.test.runner.AndroidJUnitRunner ``` The following device instrumentation evidence is a prior baseline, not a blanket device-validation claim for schema-v8 definitions or three-mode sync. That suite has **5 tests**: two real-SQLite repository checks and three production-screen Compose UI checks. Coverage includes exact raw form restoration through Activity recreation, cancellation and confirmation of discard, final save with no stale Resume after recreation, and refusal of a second completion. All tests use isolated databases; the debug-only, non-exported Activity renders production Home/Session screens without exporting synthetic data. The phone must be unlocked and interactive. Do not interpret a locked-screen `No compose hierarchies found` failure as a passing UI check. The Samsung SM_G990B additionally passed the normal `MainActivity` matrix: - two real catalog exercises with distinct continuous values survived Home, another app, background `am kill` with verified PID exit, and force-stop; - rotation recreated the Activity and Home Resume restored both exercises; - removal of one exercise survived process death, with the other intact; - raw duration/speed text survived process death and force-stop exactly; - Back and the normal new-session action preserved the existing draft; - discard cancellation preserved the DB exactly; confirmation and relaunch left all draft tables empty and no Resume action; - completed History and mobile export excluded the populated draft; - every original domain row survived migration and the entire device matrix. Final-save UI and repeated-finalization tests ran on-device with isolated data; no fictitious completed session was added to the user's history. Real migration and final state passed SQLite integrity/foreign-key checks. The pre-upgrade DB and preferences backup is outside the repository. No uninstall, package-data clear, desktop schema change or frozen artifact change is part of this repair. ## 14. Current real-data reconciliation validation Before mutation, coherent desktop and Android v8 copies were created and passed `PRAGMA integrity_check` plus `PRAGMA foreign_key_check`. The newest real Android V2 artifacts were selected, including scoped-storage `(N)` siblings. The production tools then ran in protocol order on the desktop copy: ```text equipment definitions -> exercise/session/body V2 -> equipment associations -> PC definitions/catalog/mobile/associations export -> second import ``` The result retained the desktop Marche identity with catalog `data_fields = 3`, both older speed-only Marche occurrences, the Android speed-and-distance occurrence, all original `entry_id`/positions, custom equipment, set weights and body observations. The retired Android Marche ID had no remaining desktop catalog/reference owner. The second replay changed no row counts and both integrity checks remained clean. This exercises real artifacts and production import/export code on copies. It does not substitute for a final direct-MTP run when USB ownership or sandbox permissions prevent libmtp access. `tests/test_equipment_associations_exchange.py` additionally covers the bounded v13 stale-companion bridge for the five frozen machine-exercise splits. Plate Loaded Leg Press and Treadmill pass only with matching current mobile V3 proof; generic Marche, the durable Marche+treadmill draft, Seated Leg, and Rear Delt NULL provenance remain unchanged; an unrelated exercise mismatch still fails; and replay is idempotent. For this checkpoint, the USB probe discovered the connected Samsung interface but libmtp failed at `libusb_open()`. The canonical desktop DB was also read-only to the sandbox; the first attempted definitions import failed before mutation, and a repeated dump hash plus integrity/FK checks proved it unchanged. The Android package remained force-stopped and ADB was used only for read-only inspection/pull. Applying the validated state to both real stores and running the two transport directions remains an out-of-sandbox hardware validation. Detailed retained evidence: [Android draft execution record](reviews/android_session_draft_v1_resume.md). ## 15. Training knowledge V1 validation checkpoint The current implementation passes the training-knowledge catalog validator, eight Python generation/catalog regressions, 42 desktop Meson tests, and standalone C17 public-header checks for `training_knowledge.h`, `training_context.h` and `database.h`. A targeted AddressSanitizer/Undefined- Behavior-Sanitizer build passed the two new desktop API tests. Android `testDebugUnitTest assembleDebug` passed with 56 tests, zero failures/errors, and one skipped real-v9 fixture test because `TRAINLOG_ANDROID_V9_FIXTURE` was not available. The TUI knowledge screen's UTF-8 cell-aware scrolling was tested at the 72x20 minimum terminal. The catalog validator, JSON validator and import-contract validator also pass. `git diff --check` passes. Scientific-review metadata and catalog hashes were verified separately. Temporal regressions now cover the original `+14:30`, `+15:00` and lowercase-`t` failure, lowercase `z`, omitted seconds, high/negative offsets, exact fractions, stable identity ties, cursor reuse, traversal through exhaustion and malformed stored values. C tests also cover selected/unselected values beyond its fixed output capacity. Four Python tests cover the explicit grammar and exact chronology, including actual document admission with missing, strict and permissive optional JSON Schema format checkers. An independent twelve-form production C probe passes normally and under ASan/UBSan. Fresh final validation passed: strict build; 42 Meson tests; eight Python knowledge tests; four Python temporal tests; knowledge, JSON, and import validators; three strict C17 headers; affected C knowledge/context tests under ASan/UBSan plus Python timestamp validation; normal and sanitized independent 12-form temporal probes; Android 56 tests with zero failures/errors and one known missing real-v9 fixture skip; Java 17 debug assembly; skill validation; and `git diff --check`. Generated C is byte-identical with SHA-256 `e8c099f67eb111d61621b5d76592c049823af5508d43e73ec646f22e4c377fca`; all six Android assets are byte-identical. `TRAINING_KNOWLEDGE_V1=PASS`. The independent temporal review passed with no findings; its coverage included parser grammar and limits, exact chronology, ties, cursor semantics, source capacity, snapshots, and Android/Python parity. The initial full audit's stale temporal documentation, Android loader, Meson dependency, and C role-only query findings were repaired and independently verified. See the [temporal contract](reviews/training_knowledge_v1_temporal_contract.md) for the established contract. No real Android install, manual TUI visual exercise, or manual MTP hardware validation was performed for Training Knowledge V1. Shell/naming closeout coverage includes real-PTY shell behavior, canonical name export/import replay on temporary databases, and preservation of sets, MAX, equipment, BODY ZONES, aliases, and distinct unmapped exercises. EQUIPMENT_KNOWLEDGE_V2 validation covers relation envelopes, canonical UUID and equipment references, enum/source integrity, uniqueness and stable ordering, the unresolved Seated Leg invariant, generic/custom non-inference, and generated C/Android query parity for leg press, combo equipment, and chest-zone derivation. `training_feedback_sync` covers first import, identical replay, post-sync follow-up addition, PC export convergence, and immutable-ID conflict. Android unit tests cover partial/final recognition, correction, resume/append, stop, save/cancel, failure preservation, permission/unavailable fallback, lifecycle destroy, and H+ flooring/unknown anchors. Draft-feedback coverage saves before completion, reopens the repository, preserves multiple stable IDs and ordering across ordinary draft rewrites, verifies active/completed Compose readback, excludes the draft rows from sync, transfers them atomically on finalization, rolls back on an injected insert failure, and checks discard cascading. The real-v13-copy production migration test also requires the existing active draft and all its occurrences to remain identical while the new draft-feedback table starts empty and a second production open is logically idempotent. `feedback_correction` reconstructs completed occurrence rowids while retaining, reordering and changing the exercise context of stable entries; it verifies exact feedback preservation, selective cascade on removal, no inheritance by a new entry, unchanged follow-ups, and full rollback on injected reattachment failure. Android additionally covers its bounded resumed-MAX replacement path. `tests/test_profile_state_sync.py` covers the captured Planche droite sol legacy `SETS+REPS`/`SETS+DURATION` repair, idempotent replay, direct-descendant and multi-generation stale-ancestor convergence, sibling conflict, the explicit 32-record resource bound, exact JSON scalar types, immutable machine conflict, and preservation of the old occurrence snapshot. Android `ExerciseProfileStateSyncTest` covers the same transitive/type contract in the reverse PC→Android path and future-occurrence snapshot. `HistoricalSessionProfileImportTest` exercises the production V2/V3 repository import with equal and evolved tracking/recording/data-field profiles, strict snapshot-local payload rejection, unknown identity rejection, unchanged historical values, unchanged current profile history, and idempotent replay. `SyncBundlePublicationTest` exercises the production request coordinator and a synthetic direct exchange directory. The exact BODY ZONES file is rewritten beside 31 preserved historical copies, no `(32)` copy occurs, all seven Android-origin companions precede the request signal, and replay reuses canonical names without stable-feedback churn. It also requires one direct-child enumeration for the entire export-plus-request transaction, one additional enumeration for each later synchronization, index update after a create without rescan, one write per canonical artifact, and non-main dispatcher ownership for snapshot and stream operations. Focused publisher cases additionally cover an empty directory, exact rewrite, ten successive publications without numbered copies, historical numbered copies alone and beside a canonical object, exact final content, and an explicit canonical-name conflict without writing a wrongly named target. `DirectExchangeStorageTest` verifies the manifest permission and settings action, refusal without an authorized backend, directory creation, real temporary-file replacement and truncation, ten publications without suffixes, and preservation of legacy Download and recovery fixtures. It also reconstructs a truly fresh Android database from catalog, alias, profile-state and body-zone companions, including an old zone `exercise_id` resolved through the durable alias. A deliberately malformed downstream session companion proves zones are already durable before the error; the repaired pass imports the one-sync AI draft, and replay preserves three canonical identities and exactly seven zone rows without duplicates. `AiSessionDraftUiWiringTest` additionally requires the delete action to expose the titled confirmation before mutation, verifies Cancel and Back dismissal, then proves that only confirmation removes the pending proposal while retaining its replay-blocking deleted tombstone. Its existing Start coverage remains. Native `sync_body_zone_wiring` verifies the profile pre-pass before mobile import, strict post-pass gating, replay, legacy peer compatibility, and the explicit missing-companion diagnostic even when a stale local temporary profile artifact exists. It also receives a valid alias V1 companion through the production MTP orchestration, imports it into an isolated database through the production Python argv path, and proves that the sync continues. Negative cases cover an unresolved helper, argparse failure, Python exception/stderr propagation, sync-history visibility, and rejection of a stale result-file `PASS`; the successful flow exercises the other shared Python helper callers as well. Every Android outbound artifact and the request use the shared direct-storage exact-name resolver; app-scoped MediaStore visibility is irrelevant. `profile_state_sync` includes the production persisted update-trigger dependency on `trainlog_profile_revision`, imports through the same Python connection path used by the PC sync worker, and requires first import plus identical replay to pass without revision-state or history-count churn. It also checks that the central factory installs both profile-revision schema functions. `TRAINLOG_ANDROID_DIRECT_STORAGE_V1` coverage fixes Android storage at `/storage/emulated/0/Documents/Trainlog`, preserves canonical rewrite/no-suffix behavior, ignores legacy Download trees, and gives the MTP fixture both Documents and Download roots while failing on any legacy access. Android unit tests and any instrumented tests run on an emulator; the primary personal phone is reserved for manual, non-destructive install-and-sync smoke validation and must not run `connectedDebugAndroidTest`. ## Generation, transaction, and acknowledgement V1 The normal Meson inventory includes `sync_generation_exchange` and `schema_v21_migration`. The service suite exercises deterministic desktop snapshot capture against a controlled writer, immutable publication/replay, partial and substituted input, strict paths/capabilities/bounds, wrong ACK context, durable ACK replay after reopen, late-domain rollback, rejected ACKs, and unchanged causal-operation IDs/digests across first emission and retransmission for all seven causal target kinds. The v20 fixture migration preserves populated business state and creates no historical generation/ACK. Android `SyncGenerationServiceTest` uses production repository codecs and the filesystem object adapter. It proves a transaction-held capture against a deterministically blocked writer, v19→v20 preservation, manifest-last immutable publication, whole-generation rollback on a late semantic failure, committed ACK replay after close/reopen, and causal-operation byte stability. Its real cross-implementation test runs both chains: an Android-created session is captured and transactionally consumed by the desktop Python service, and a separate desktop-created session written through the public C17 database API is captured by the desktop service and consumed by Android. Both ACKs return to the original producer. The bridge moves bytes and invokes real entry points; it does not author expected business JSON or repair artifacts. After these additions the direct normal suites report 78/78 Meson tests and 216 Android tests: 212 passed, four historical-fixture skips, zero failures. The grouped Web end-to-end tranche raises the direct normal and sanitizer inventories to 79/79 and adds `sync_orchestrator`, protected sync API coverage, frontend sync assertions and a real headless Firefox scenario in `tests/test_web_sync_browser.py`. Its detailed matrix is [`sync_web_end_to_end_v1_evidence.md`](design/sync_web_end_to_end_v1_evidence.md). The browser uses the production C server and worker with an isolated directory transport fixture; physical MTP and Drive are not claimed. The publication adapter is an isolated directory/object-store double. No physical MTP, phone, Drive, active service, or user database participates. ## Causal deletion V1 The normal Meson inventory includes `causal_delete_exchange`, which invokes the production desktop service and proves all target kinds, relationship safety, exact replay, conflicts, finalization protection, and rollback. Android `causalSessionDeletionRoundTripsWithBothRealProducersAndBlocksReplay` uses the real Android repository and desktop service in both producer directions with fresh destinations, reopen/replay, and legacy non-resurrection. `causalFiveDomainDeletionsRoundTripWithAndroidAndDesktopProducers` closes the remaining producer/consumer matrix. The Android repository creates exercises, populated observations, custom equipment, multi-revision feedback and BODY ZONE relations; production V4 and companion tools establish the desktop replica. Android-local deletion is consumed and replayed by the desktop service. A second production-provisioned desktop replica creates each deletion locally, exports it to Android, and receives Android's exact replay. | Kind | Producer -> consumer | Stored/read-model assertion | Result | |---|---|---|---| | `exercise` | Android -> desktop | Historical row/reference retained; new-work predicate excludes it; unrelated exercise and alias protection remain | PASS | | `exercise` | desktop -> Android | Repository list excludes it after reopen; history row remains; stale V4/catalog state cannot restore it | PASS | | `body_observation` | Android -> desktop | Populated target row is absent; session and unrelated populated observation remain | PASS | | `body_observation` | desktop -> Android | Repository list excludes the target after reopen; unrelated observation and session remain | PASS | | `custom_equipment` | Android -> desktop | Custom definition/history remains but new-work predicate excludes it; built-in occurrence equipment remains | PASS | | `custom_equipment` | desktop -> Android | Historical equipment row/reference remains; repository list excludes it; built-in equipment remains | PASS | | `feedback` | Android -> desktop | Root and two immutable revisions remain; current predicate excludes it; unrelated feedback remains | PASS | | `feedback` | desktop -> Android | Root/revisions remain after reopen; repository current view excludes it; unrelated feedback remains | PASS | | `body_zone_relation` | Android -> desktop | Target secondary relation is absent; primary relation and exercise remain | PASS | | `body_zone_relation` | desktop -> Android | Target secondary relation is absent after reopen; primary relation and exercise remain | PASS | Every row uses a legitimate replica established through production exchange entry points. Repeated local deletion returns the original immutable operation; the opposite exporter preserves its operation ID, target, predecessor and payload; replay after a new database connection is unchanged. Protected V4, equipment, feedback and BODY ZONES legacy imports refuse stale restoration and leave all accepted rows and causal state unchanged. Generic codec reuse is not counted as a domain result. Strict compilation uses Clang 22.1.8. GCC 16.2.1 retains the compared pre-existing `-Werror=misleading-indentation` failure in unchanged Web baseline sources; it is not reported as green and warning policy is unchanged. The causal closeout full harness result is 76/76 normal Meson, 76/76 ASan/UBSan, and 211 Android tests: 207 passed and four optional historical database fixtures skipped. `assembleDebug`, JSON validation, import-contract validation, both frontend tests, and the isolated characterization all pass. Focused closeout assertions are `test_child_mutations_invalidate_session_observation_and_feedback_deletes`, `test_imported_builtin_and_alias_retirement_are_refused_without_writes`, `test_draft_identity_is_semantic_revision_not_json_serialization`, and `test_operation_count_and_input_byte_bounds_precede_effects` in the desktop production-service suite. Android adds `causalDraftDeletionUsesTheSameRevisionForActivePendingAndDesktopReplicas` and `importedCausalOperationsCannotRetireBuiltInIdentities`; the existing session round trip now also changes and reverts a confirmed set before proving the remote deletion conflicts without losing business or causal state. ## Generation MTP and Android backup `generation_mtp` drives the production transport algorithm through typed fake libudev/libmtp callbacks. It covers exact peer selection, bounded download, manifest-last publication, ambiguity and truncated-object rejection. The Firefox end-to-end scenario also runs in `mtp` mode: only the callback table is replaced by a filesystem object store while the C transport, Python worker, Android generation service, desktop consumers and durable ACK path remain real. `AndroidBackupServiceTest` creates and restores a populated current-schema backup including an active draft, raw partial input, stable peer identity and preferences, and rejects corrupt and traversal entries. The bridge build runs the analogous test on the exact schema-v17 source and feeds its archive to `BridgeBackupMigrationTest`, which verifies normal migration to schema v21. These are software proofs. They do not claim physical USB/MTP behavior, installation over the signed release, a real user backup, or a real-device restore. The production Android coordinator and desktop peer worker also complete 24 successive bidirectional conversations against their real SQLite services with no cleanup between runs. The test crosses multiple archive cycles and checks that the active draft survives. After the eighth exchange it also reproduces the deployed legacy evidence gap by retaining acknowledged generations while removing only their producer-side ACK rows; the next production conversation must restore the desktop consumer's exact correlated ACKs, archive eligible generations, and complete without deleting or duplicating any generation. Focused archive tests cover interrupted-copy resume, invalid archives, simulated insufficient space, missing and late ACKs, and idempotent ACK replay after archival. The controlled Samsung SM-G990B rollout migrated the authentic Android backup from schema v20 to v21 and an authentic desktop copy from v21 to v22 before installation. The installed release retained its existing certificate and advanced from versionCode 6 through 7 to 8 without uninstalling or clearing data. Real MTP validation preserved the failed-run evidence that exposed an Android external-storage directory-fsync incompatibility and an early desktop generation reference. After their bounded repairs, run `sy_06200dd7-0d51-4f33-b80d-9976822ba87d` completed, the two Trainlog services were restarted, and run `sy_783014b0-f3f0-4f1e-8f4f-723acfc02dbc` completed. Both exchanged distinct generations and durable ACKs with zero session reconciliation. The final Android backup passed integrity and foreign-key checks at schema v21 with 13 generations, five verified archives, 13 ACKs, five consumed generations, and six active generations. The expected AI proposal remained `pending` with its original identity and six entries. ## Readability validation The normalized C scope is checked non-mutatively with the explicit `clang-format --dry-run --Werror` command in `docs/coding_style.md`. Clang C17, the normal Meson inventory and the ASan/UBSan inventory preserve the existing behavioral assertions. Python syntax/orchestrator tests, Android generation and backup tests, the I/O-boundary MTP browser proof, JSON/import validators and the isolated harness cover the other normalized sources. GCC 16.2.1 and Clang 22.1.8 both pass separate strict C17 builds and the full **81/81** Meson inventory with `werror=true` and warning level 3. The former GCC `format-truncation` findings are closed by exact checked path construction in the generation-MTP tests/double and by a production accepted-response serializer that reports its byte count. Boundary tests cover exact fit, one-byte insufficiency and multibyte path components; failure leaves no partial path or response available to a caller. Clang ASan/UBSan passes the same **81/81** inventory with leak detection. The readability closeout preserves the **81/81** normal Meson and **81/81** ASan/UBSan inventories. The isolated Android run reports **220 tests: 215 passed, five skipped, zero failures/errors**, and `assembleDebug` succeeds on JDK 17. The real Firefox/object-I/O-boundary MTP conversation also passes after formatting. A source comparison against the pre-readability commit confirms that the scoped test names and assertion counts are unchanged. The deployment-tools regression also constructs the relocatable candidate and requires its inventory to include the packaged `trainlog-sync-once` binary, the `trainlog-syncd` launcher and daemon source, and the generation worker. This prevents a user unit from silently executing a helper from a different checkout than the installed desktop candidate. `syncd_routing` proves that both Android-labelled and daemon-labelled admissions invoke the packaged `sync_orchestrator.py` full-generation owner when trusted configuration is present, and never invoke the legacy `trainlog-sync-once` command in that mode. The generation-MTP boundary test proves that polling copies the Android request signal without acknowledging or deleting it. Session-exchange characterization keeps standalone V3 rejection strict while its complete-envelope case filters a tombstone-dominated live fact; the causal Android/Desktop round-trip and generation replay/ACK tests cover restart, idempotence and non-resurrection. `AndroidBackupServiceTest` injects a product version and verifies that exact value in the ZIP manifest, preventing a return to a hard-coded release string. The isolated JDK 17 harness reports **220 Android tests: 215 passed, five skipped, zero failures/errors**. The skips are optional, externally supplied copy gates: `BridgeBackupMigrationTest.verifiedSchemaSeventeenBridgeBackupMigratesThroughCurrentOwner`, `RealAndroidV9BodyZonesMigrationTest.realVersionNineCopyMigratesWithoutChangingExistingTables`, `RealAndroidV14FeedbackRevisionMigrationTest.freshRealCopyMigratesLosslesslyToVersionFifteen`, and the version-thirteen and version-twelve methods in `RealAndroidV12MachineExerciseMigrationTest`. They require respectively the bridge backup input, a real v9 fixture, a real v14 fixture/output directory, and real v13/v12 fixtures with explicit output directories. Synthetic bridge and current-schema migration proofs run separately; no user database is read. The bridge migration test accepts both legitimate backup states: an active draft may be present or absent. It snapshots every original table, column, SQLite value type and value before restore, migrates through the production repository owner, and compares that complete historical shape afterward. The draft assertion follows the archived state instead of requiring the synthetic fixture's sample draft. ## Private full-generation rollout validation The authorized 2026-09-17 rollout added regression coverage for production failures found only after installation: feedback import under `sqlite3.Row`, the complete exporter inventory of the relocatable package, stale Android ACK and generation-reference correlation, and the completed-generation UI state. The real direct-libmtp smoke then completed three bidirectional conversations on the paired phone. Restart and replay preserved identical logical hashes for 23 desktop business tables and 540 rows; SQLite integrity and foreign-key checks passed, both outgoing desktop generations were acknowledged, and the Android UI displayed completion. The private journal retains exact IDs and hashes. Drive and `connectedDebugAndroidTest` were not run; the personal phone was used only for non-destructive install, backup and manual sync validation.