diff --git a/docs/design-plans/2026-08-07-remove-vault-sync.md b/docs/design-plans/2026-08-07-remove-vault-sync.md
new file mode 100644
index 00000000..522f5494
--- /dev/null
+++ b/docs/design-plans/2026-08-07-remove-vault-sync.md
@@ -0,0 +1,426 @@
+# Remove Vault Sync Design
+
+## Summary
+
+This design deletes the app's Obsidian vault-sync integration: the HTTP bridge client, the offline
+sync queue, the settings fields and screens that configure it, and the `sync_status` tracking that
+gates routine deletion. It also strips vault references from user-facing copy and from AGENTS.md.
+What survives on purpose is the markdown serialization layer (`src/interop`) and the unused export
+scaffolding (`src/export`) — both are left running and green so a future Excel-based backup path
+can build on them without redoing the parsing and serialization work.
+
+The approach is four independently-mergeable phases, each leaving `npm test` and `tsc --noEmit`
+green: first de-wire every caller of `src/sync` (including a **dynamic**
+`import('@/sync/syncService')` inside `onCompleteSession` that the compiler cannot see, making this
+phase hand-audited rather than compiler-driven); then delete the sync modules and settings fields
+outright; then drop the `sync_status` column from the WatermelonDB schema — via WatermelonDB's
+undeclare-rather-than-drop convention, since column deletion is not a supported migration step —
+and remove the delete-guard it existed to serve, making every routine deletable; and finally sweep
+user-facing copy and documentation so nothing still describes a bridge that no longer exists. Two
+responsibilities the sync layer quietly duplicated, a zero-set defaulting rule and a
+null-normalization step, become solitary rather than redundant once it is gone, and the plan calls
+out exactly where the remaining single owner needs its test coverage confirmed before the redundant
+copy disappears.
+
+## Definition of Done
+
+1. **The HTTP bridge is gone from the app.** `src/sync/` (bridgeClient, syncService, and their four test files), `src/helpers/settingsActions.ts`, and the `baseUrl`/`token` fields on `BridgeSettings` are deleted. No code path posts to or reads from the Mac-side bridge.
+
+2. **No UI offers or mentions vault sync.** The `settings/bridge.tsx` screen and its Settings-index row are deleted, the Routines tab's "Import Routines" button is deleted, and every user-facing string referencing the vault or syncing is removed or reworded — the session-abandon confirmation, the Today tab empty state, and the delete-routine error.
+
+3. **`sessions.sync_status` is dropped at schema v4.** The column is removed from the `appSchema` declaration, not from existing SQLite tables — WatermelonDB 0.28 has no supported column-removal step (see Architecture). `Session.customSyncStatus`, `deleteRoutine`'s unsynced-session guard, and `RoutineHasUnsyncedSessionsError` are all removed — routines become always-deletable.
+
+4. **`src/interop` and `src/export` survive untouched**, still compiling and still green, so a future export path (Excel or otherwise) can build on them.
+
+5. **AGENTS.md drops the two-repo split, the Sync section, and the contract-mirroring rule.** `../workout-bridge` itself is left alone on disk.
+
+6. **`npm test` and `npx tsc --noEmit` pass**, and a session completes end-to-end in the simulator with no sync errors.
+
+**Out of scope:** Excel backup; wiring `src/export` to any UI; any replacement routine-authoring path (the AI Coach becomes the sole way to create routines, and that is accepted).
+
+## Acceptance Criteria
+
+### remove-vault-sync.AC1: The bridge is unreachable from the app
+- **remove-vault-sync.AC1.1 Success:** Completing a session issues no network request to a bridge
+  URL — the effect path writes the database and HealthKit only.
+- **remove-vault-sync.AC1.2 Success:** `src/sync/` does not exist, and no file under `src/` imports
+  `@/sync/…` either statically or dynamically.
+- **remove-vault-sync.AC1.3 Success:** The settings type has no `baseUrl` or `token` field, and
+  `getSettings()` returns an object without them.
+- **remove-vault-sync.AC1.4 Edge:** A stored settings blob that still contains `baseUrl`/`token`
+  loads without error, and those stale keys reach no read site.
+- **remove-vault-sync.AC1.5 Success:** `src/helpers/settingsActions.ts` and its test no longer exist.
+
+### remove-vault-sync.AC2: No UI offers or mentions vault sync
+- **remove-vault-sync.AC2.1 Success:** The Settings index renders exactly one row ("AI") and offers
+  no route to a bridge screen.
+- **remove-vault-sync.AC2.2 Success:** `/settings/bridge` resolves to no screen.
+- **remove-vault-sync.AC2.3 Success:** The Routines tab renders the AI Coach button and no "Import
+  Routines" button.
+- **remove-vault-sync.AC2.4 Success:** Deleting a routine that has a completed session succeeds with
+  no error alert.
+- **remove-vault-sync.AC2.5 Success:** The session-abandon confirmation conveys that the deletion is
+  permanent without referencing a vault.
+- **remove-vault-sync.AC2.6 Success:** The Today tab's no-routines empty state names the AI Coach as
+  the only way to create a routine and does not mention import or the vault.
+- **remove-vault-sync.AC2.7 Edge:** No user-visible string in `src/app/` or `src/components/`
+  contains "vault", "sync", or "bridge".
+
+### remove-vault-sync.AC3: `sync_status` is gone from the schema and the code
+- **remove-vault-sync.AC3.1 Success:** `src/db/schema.ts` declares `version: 4` and the `sessions`
+  table has no `sync_status` column.
+- **remove-vault-sync.AC3.2 Success:** `src/db/migrations.ts` contains a `toVersion: 4` entry that
+  `schemaMigrations` accepts with an empty `steps` array.
+- **remove-vault-sync.AC3.3 Success:** A database created fresh at v4 has no `sync_status` column,
+  and `createSession` writes a session without one.
+- **remove-vault-sync.AC3.4 Success:** The `Session` model has no `customSyncStatus` field.
+- **remove-vault-sync.AC3.5 Success:** `deleteRoutine` deletes a routine referenced by a completed
+  session, and `RoutineHasUnsyncedSessionsError` is no longer exported from `src/db/repository.ts`.
+- **remove-vault-sync.AC3.6 Edge:** An existing v3 database opens at v4 without throwing. Data
+  preservation across the upgrade is explicitly not required (pre-release app).
+
+### remove-vault-sync.AC4: The markdown layer survives intact
+- **remove-vault-sync.AC4.1 Success:** `src/interop/` and `src/export/` exist unchanged and their
+  test suites pass.
+- **remove-vault-sync.AC4.2 Success:** `serializeRoutine`, `serializeSession`, `parseRoutine`, and
+  `parseSession` remain exported and importable.
+- **remove-vault-sync.AC4.3 Success:** `upsertRoutine` still defaults `targetSets` to 1 for an entry
+  with absent `targetSets` and zero `warmupSets`, proven by a test that does not depend on `src/sync`.
+- **remove-vault-sync.AC4.4 Failure:** `validateRoutineDraft` still rejects a draft with an explicit
+  `targetSets: 0`.
+
+### remove-vault-sync.AC5: Documentation matches the code
+- **remove-vault-sync.AC5.1:** AGENTS.md contains no "Two-repo split" section, no "Sync (`src/sync`)"
+  section, and no `src/sync/` entry under Structure.
+- **remove-vault-sync.AC5.2:** AGENTS.md's Boundaries section no longer requires mirroring markdown
+  contract changes into `../workout-bridge/src/contract.ts`.
+- **remove-vault-sync.AC5.3:** AGENTS.md's interop section names `buildSessionSetLine`'s `!= null`
+  guards as the sole null-normalization layer and no longer instructs "keep both layers."
+- **remove-vault-sync.AC5.4:** AGENTS.md's zero-total defaulting passage names `upsertRoutine` as the
+  only enforcing layer and drops the "redundant by construction" framing.
+- **remove-vault-sync.AC5.5:** Engine convention 1 no longer cites vault sync as a driver of
+  `CompleteSession`.
+- **remove-vault-sync.AC5.6:** `jest.config.js`'s `testMatch` no longer lists `sync`.
+- **remove-vault-sync.AC5.7:** `src/db/repository.ts`'s `deleteSession` and `deleteRoutine` comments
+  no longer explain their semantics in terms of `syncNow()`.
+- **remove-vault-sync.AC5.8:** AGENTS.md's "Last verified" date is updated.
+
+### remove-vault-sync.AC6: Cross-cutting behaviors
+- **remove-vault-sync.AC6.1:** `npm test` passes at every phase boundary.
+- **remove-vault-sync.AC6.2:** `npx tsc --noEmit` reports no errors at every phase boundary.
+- **remove-vault-sync.AC6.3:** `npm run lint` passes at every phase boundary.
+- **remove-vault-sync.AC6.4:** In the simulator, a routine starts, sets log, the session completes
+  with no sync-related console error, and the session appears in History.
+- **remove-vault-sync.AC6.5:** In the simulator, Settings → AI opens and saves an API key, proving
+  the `bridge_settings` storage key still resolves after the type rename.
+
+## Glossary
+
+- **Vault**: The user's Obsidian markdown vault — the external system this feature synced workout
+  data to and from. Vault sync is the entire feature being removed.
+- **Bridge (`../workout-bridge`)**: A separate Mac-side Node/vitest HTTP service, reached over
+  Tailscale, that read and wrote the vault's `_sync/` folder on the app's behalf. This design
+  deletes the app's client for it but leaves the bridge repo itself untouched.
+- **WatermelonDB**: The on-device database library (SQLite-backed on iOS, LokiJS on web) that
+  stores routines and sessions. Its schema/migration system is why `sync_status` is *undeclared*
+  rather than dropped — WatermelonDB 0.28 has no supported "remove a column" migration step.
+- **LokiJS**: The in-memory JavaScript database WatermelonDB uses as its web-platform adapter, in
+  place of SQLite. Mentioned because it silently ignores raw-SQL migration steps, which is part of
+  why `unsafeExecuteSql` was rejected as an option.
+- **Schema migration (`schemaMigrations`, `toVersion`, `steps`)**: WatermelonDB's versioned
+  mechanism for evolving the on-device database shape over app updates. Each release bumps
+  `version` and adds a step describing the change; this design adds a v4 entry with deliberately
+  empty steps.
+- **`sync_status`**: The database column (on `sessions`) that tracked whether a completed session
+  had been posted to the vault yet (`'local'` vs `'synced'`). Its removal is what makes every
+  routine deletable, since the old delete-guard existed only to protect unsynced sessions.
+- **Dynamic import**: A JavaScript/TypeScript `await import(...)` call resolved at runtime rather
+  than compile time. Flagged here because `onCompleteSession` reaches the sync module this way, so
+  `tsc` cannot catch a broken reference to it — the reason Phase 1 is hand-audited rather than
+  compiler-verified.
+- **`onCompleteSession`**: The shell-side function that runs when a workout session finishes,
+  currently triggering the sync post, the HealthKit export, and a post-workout AI debrief in
+  sequence. The design requires the sync step be removed without disturbing the order of the other
+  two.
+- **Markdown contract (`src/interop`)**: The app's shared serialization and parsing logic for the
+  vault's markdown format (routines and sessions). It survives this change untouched, since it is
+  the reusable piece the sync layer was built on top of, not the sync layer itself.
+- **AI Coach**: The app's conversational, Anthropic-API-backed feature for authoring workout
+  routines. It becomes the *only* way to create a routine once vault import is deleted.
+- **HealthKit**: Apple's health-data framework, which the app writes completed workouts to
+  (write-only, no reads). Unaffected by this change; called out only to confirm it is not reordered
+  relative to the deleted sync step.
+- **Zero-total defaulting**: A validation rule ensuring a routine entry that specifies only a
+  duration (no explicit set count) still gets `targetSets` defaulted to 1, so it is not silently
+  unstartable. Two copies of this rule currently exist (one in the deleted sync layer, one in
+  `upsertRoutine`); the design requires confirming the surviving copy is actually tested before the
+  other is deleted.
+- **`testMatch` (jest.config.js)**: The glob pattern controlling which `src/` subdirectories jest
+  treats as tests. Removing `sync` from this list is part of the deletion, since no code is left in
+  that directory to test.
+
+## Architecture
+
+This is a deletion, not a construction. The architecture question is not what to build but what
+each removal leaves behind, and which surviving code silently inherits a responsibility it did
+not have before.
+
+### What is removed
+
+The Mac-side HTTP bridge is the *transport* layer of the vault integration. It comprises
+`src/sync/bridgeClient.ts` (HTTP, with `BridgeUnreachable` vs `BridgeHttpError` failure types),
+`src/sync/syncService.ts` (the offline queue: `syncNow` posts `sync_status='local'` sessions and
+flips them to `'synced'`; `importRoutines` pulls vault routines), `src/helpers/settingsActions.ts`
+(UI-facing wrappers), the `settings/bridge.tsx` screen, and the `baseUrl`/`token` settings fields.
+
+### What survives, and why
+
+`src/interop/` (the markdown contract: `format.ts`, `parse.ts`, `serialize.ts`) and `src/export/`
+(`exportService.ts`, `exportPresenter.ts`) are untouched. `src/export` is already wired to no UI —
+it was built under `docs/design-plans/2026-08-04-vault-export.md`, whose stated plan was that
+bridge removal happens in a follow-up ticket. This is that ticket. Keeping both modules preserves
+a working markdown serializer for the Excel backup work to build on or reference, at the cost of
+carrying dead code. After Phase 2, `parse.ts` has no production consumer at all; only tests
+(`src/export/exportService.test.ts`, `src/state/startSessionFromRoutine.integration.test.ts`)
+import it. That is accepted, not an oversight.
+
+### Three inherited responsibilities
+
+Deleting `syncService.ts` transfers three obligations to code that currently shares them:
+
+1. **Zero-total defaulting.** `defaultTargetSetsForDurationLine` (`src/sync/syncService.ts:24`)
+   defaults a duration-based routine entry's `targetSets` to 1 when it is absent and `warmupSets`
+   is 0, preventing an entry that plans zero sets from reaching the engine. `upsertRoutine`
+   (`src/db/repository.ts`) applies the identical rule. AGENTS.md documents the sync-layer copy as
+   "redundant by construction" because `upsertRoutine` sees every write either layer produces.
+   That redundancy is what makes the deletion safe — but it also means the surviving layer becomes
+   solitary, and the tests proving the rule (`src/sync/syncService.test.ts:732`) are the ones being
+   deleted. Phase 2 must confirm `upsertRoutine`'s own tests cover the case before removing layer 1.
+
+2. **Null normalization at the shell boundary.** `syncService.ts` maps DB rows to the serializer
+   with `?? undefined`, because WatermelonDB returns `null` for unset optional columns while
+   `buildSessionSetLine` guards on `!= null`. AGENTS.md instructs "keep both layers." After
+   removal there is one layer, and the instruction is wrong as written.
+
+3. **Session-completion side effects.** `onCompleteSession` (`src/state/activeSession.ts:207-233`)
+   reaches sync through a **dynamic** `await import('@/sync/syncService')`. TypeScript cannot see
+   through a dynamic import, so deleting the module produces a runtime failure at session
+   completion rather than a compile error. This is the single highest-risk edit in the change and
+   the reason Phase 1 exists as a separate, hand-audited phase rather than a compile-error sweep.
+
+### Schema strategy
+
+WatermelonDB 0.28.0 exports exactly four migration step builders — `schemaMigrations`,
+`createTable`, `addColumns`, `unsafeExecuteSql` — verified against the installed
+`Schema/migrations/index.d.ts`. The [official documentation](https://watermelondb.dev/docs/Advanced/Migrations)
+states that deleting columns is "not yet implemented"; `destroyColumn` was merged upstream in
+PR #1799 but has not shipped in a release. Official guidance is to leave unused columns in the
+database and omit them from the schema definition, which WatermelonDB then ignores during CRUD.
+
+The design therefore **undeclares** rather than drops:
+
+- `src/db/schema.ts` — `version: 3 → 4`, `sessions.sync_status` column declaration deleted.
+- `src/db/migrations.ts` — `{ toVersion: 4, steps: [] }`. An empty `steps` array passes
+  `schemaMigrations`' validation (`Array.isArray(steps) && steps.every(...)`), verified against
+  the installed source. The entry exists so `version` stays honest with the declaration and the
+  migration list has no gap, which `schemaMigrations` enforces.
+
+`unsafeExecuteSql('ALTER TABLE sessions DROP COLUMN sync_status;')` was considered and rejected:
+transaction guarantees for raw SQL in migrations are undocumented (Nozbe/WatermelonDB issue #1835
+is unanswered by maintainers), and the LokiJS adapter used by `adapter.web.ts` ignores SQL steps
+entirely, so the two platforms would diverge. The gain would be cosmetic.
+
+The app is pre-release, so preserving on-device data across the upgrade is explicitly not a
+requirement. This removes the only high-consequence failure mode the migration would otherwise
+carry.
+
+### Settings storage
+
+`src/state/settings.ts` persists a single JSON blob under the key `'bridge_settings'`. That blob
+also holds `anthropicKey`, `openaiKey`, `aiGoals`, `aiEquipment`, `aiPersonality`, `profileAge`,
+`profileExperience`, and `onboardingState`. **The storage key must not be renamed** — doing so
+orphans every existing user's API key and resets onboarding. The TypeScript type name
+`BridgeSettings` is already a misnomer per AGENTS.md and may be renamed freely; the on-disk key
+may not.
+
+No settings migration is needed. `loadSettings` merges with `cache = { ...cache, ...parsed }`, so
+`baseUrl`/`token` keys in an existing blob survive as undeclared extras that no read site touches —
+the same undeclare-don't-delete shape as the schema decision.
+
+## Existing Patterns
+
+Investigation confirms this design follows patterns already established in the codebase:
+
+- **Phase-leaves-main-green.** Each phase is independently mergeable with `npm test` and
+  `tsc --noEmit` passing, matching how prior multi-phase work in this repo is structured.
+- **Undeclare-don't-delete for persisted data.** Schema migration v3 (`src/db/migrations.ts`)
+  added `session_sets.exercise_id` as nullable and "deliberately not backfilled," letting existing
+  rows resolve through the old path. The schema and settings decisions here are the mirror image
+  of that same principle.
+- **Injected dependencies at shell boundaries.** `onCompleteSession`'s `syncFn` parameter exists
+  purely for test injection. Removing the sync call removes the parameter; the surrounding
+  HealthKit write keeps its own `HealthKitSaveDeps` injection, which is unaffected.
+- **Test placement follows `jest.config.js` `testMatch`.** The node project globs
+  `src/{engine,db,interop,state,sync,health,helpers,ai,theme,watch,components,export}`. Removing
+  `sync` from that list is part of the change. New tests land in `db` and `state`, both already
+  covered.
+
+Divergence: none. No new pattern is introduced.
+
+## Implementation Phases
+
+<!-- START_PHASE_1 -->
+### Phase 1: De-wire callers from sync
+
+**Goal:** No code path reaches `src/sync`, while the module still exists. This isolates the
+runtime-only hazard from the deletion.
+
+**Components:**
+- `src/state/activeSession.ts` — remove the dynamic-import sync block from `onCompleteSession`
+  (lines ~207-233) and the `syncFn` parameter from `createActiveSessionStore`. Hand-audited, not
+  compiler-driven: `tsc` cannot see a dynamic import. The HealthKit write and the post-workout
+  debrief hook that follow it are unaffected and must remain in their current order.
+- `src/app/(tabs)/routines.tsx` — delete the "Import Routines" button, the `importLabel`/
+  `importing`/`importMessage`/`onImport` props on `EntryPointButtons`, and the `bridgeClient`/
+  `syncService`/`settingsActions` imports. The AI Coach button remains.
+- `src/app/(tabs)/settings/index.tsx` — delete the "Bridge & Sync" entry from the sections list
+  and narrow the `href` union type. The two-level settings nav (`settings/_layout.tsx`) stays as
+  it is.
+- `src/app/(tabs)/settings/bridge.tsx` — delete the file.
+- `src/state/activeSession.test.ts` and any test passing `syncFn` — update call sites.
+
+**Dependencies:** None.
+
+**Covers:** `remove-vault-sync.AC1.1`, `remove-vault-sync.AC2.1`, `remove-vault-sync.AC2.2`,
+`remove-vault-sync.AC2.3`
+
+**Done when:** `grep -rn "@/sync/" src --include=*.ts --include=*.tsx` returns matches only inside
+`src/sync/` itself; `npm test` and `tsc --noEmit` pass; a session completes in the simulator with
+no sync-related console output.
+<!-- END_PHASE_1 -->
+
+<!-- START_PHASE_2 -->
+### Phase 2: Delete the sync modules and settings fields
+
+**Goal:** Remove the bridge code and its configuration surface.
+
+**Components:**
+- Delete `src/sync/` entirely — `bridgeClient.ts`, `syncService.ts`, `bridgeClient.test.ts`,
+  `syncService.test.ts`, `syncService.integration.test.ts`, `abandonedSessionNeverSyncs.test.ts`,
+  `unsyncedSessionSurvivesRoutineEdit.test.ts`.
+- Delete `src/helpers/settingsActions.ts` and `src/helpers/settingsActions.test.ts`.
+- `src/state/settings.ts` — remove `baseUrl` and `token` from the settings interface and
+  `DEFAULT_SETTINGS`. Optionally rename the type `BridgeSettings → AppSettings` (with call-site
+  updates in `src/state/coachOnboarding.ts` and elsewhere). **Do not change the `SETTINGS_KEY`
+  value `'bridge_settings'`.**
+- `src/state/settings.test.ts` — remove `baseUrl`/`token` assertions; add a case proving a stored
+  blob containing legacy `baseUrl`/`token` keys loads without error.
+- `src/ai/contextBuilder.test.ts` and sibling prompt tests (`alternatesPrompt.test.ts`,
+  `exerciseQuestionStore.test.ts`, `restCommentaryStore.test.ts`, `exerciseReplaceStore.test.ts`) —
+  the secret-leak regression asserts `baseUrl`/`token` never reach a prompt. Narrow those
+  assertions to `anthropicKey`/`openaiKey` rather than deleting them; the guard keeps its purpose
+  and loses only two subjects.
+- `src/db/repository.test.ts` — verify `upsertRoutine`'s zero-total defaulting is covered here
+  before the `defaultTargetSetsForDurationLine` tests are deleted with `syncService.test.ts`. Add
+  the coverage if it is missing.
+- `jest.config.js` — remove `sync` from the `testMatch` glob.
+
+**Dependencies:** Phase 1.
+
+**Covers:** `remove-vault-sync.AC1.2`, `remove-vault-sync.AC1.3`, `remove-vault-sync.AC1.4`,
+`remove-vault-sync.AC1.5`, `remove-vault-sync.AC4.1`, `remove-vault-sync.AC4.2`,
+`remove-vault-sync.AC4.3`, `remove-vault-sync.AC4.4`
+
+**Done when:** `src/sync/` does not exist; `npm test` and `tsc --noEmit` pass; `src/interop` and
+`src/export` suites still pass unchanged.
+<!-- END_PHASE_2 -->
+
+<!-- START_PHASE_3 -->
+### Phase 3: Schema v4 and the always-deletable routine
+
+**Goal:** Remove `sync_status` from the schema and every read site, and drop the delete guard it
+existed to serve.
+
+**Components:**
+- `src/db/schema.ts` — `version: 4`; delete the `sync_status` column from the `sessions` table.
+- `src/db/migrations.ts` — add `{ toVersion: 4, steps: [] }` with a comment explaining that the
+  empty steps array is deliberate, that WatermelonDB 0.28 has no column-removal step, and that the
+  physical column is left in place and ignored.
+- `src/db/models/Session.ts` — delete `customSyncStatus`.
+- `src/db/repository.ts` — `createSession` stops writing `sync_status: 'local'`; delete
+  `deleteRoutine`'s unsynced-session check and the exported `RoutineHasUnsyncedSessionsError`
+  class; rewrite the `deleteSession` (~lines 258-271) and `deleteRoutine` (~lines 1254-1259)
+  comment blocks, which currently explain their semantics purely in terms of `syncNow()`'s
+  candidate selection.
+- `src/app/(tabs)/routines.tsx` — remove the `RoutineHasUnsyncedSessionsError` import and its
+  catch arm.
+- `src/db/migrations.test.ts` — add a v3→v4 case.
+- `src/db/repository.test.ts` — add a test asserting a routine referenced by a completed session
+  is deletable (the inverse of the deleted `unsyncedSessionSurvivesRoutineEdit.test.ts`).
+
+**Dependencies:** Phase 2.
+
+**Covers:** `remove-vault-sync.AC2.4`, `remove-vault-sync.AC3.1`, `remove-vault-sync.AC3.2`,
+`remove-vault-sync.AC3.3`, `remove-vault-sync.AC3.4`, `remove-vault-sync.AC3.5`,
+`remove-vault-sync.AC3.6`
+
+**Done when:** `npm test` and `tsc --noEmit` pass; a fresh simulator install starts a routine,
+logs sets, completes the session, and shows it in History; a routine with a completed session
+deletes without an alert.
+<!-- END_PHASE_3 -->
+
+<!-- START_PHASE_4 -->
+### Phase 4: User-facing copy and documentation
+
+**Goal:** No string or document asserts something the code no longer does.
+
+**Components:**
+- `src/app/session.tsx:404` — the abandon confirmation's "It will not be saved to your vault"
+  clause carried the warning's weight; the replacement must convey permanence without it.
+- `src/app/(tabs)/index.tsx:165` — the no-routines empty state drops "or import your routines from
+  the vault," leaving the AI Coach as the only named path.
+- `src/app/(tabs)/index.tsx:57-58` — the `useFocusEffect` rationale comment cites the vault import
+  as a routine source.
+- `AGENTS.md` — delete the "Two-repo split" and "Sync (`src/sync`)" sections, the `src/sync/`
+  entry under Structure, and the Boundaries rule requiring mirroring into
+  `../workout-bridge/src/contract.ts`. Rewrite three passages whose *content* changes: the interop
+  section's "keep both layers" null-normalization instruction (one layer remains); the zero-total
+  defaulting paragraph (`upsertRoutine` is now solitary, not redundant); and engine convention 1,
+  which justifies `DiscardSession` by citing vault sync alongside the HealthKit export. Update
+  "Last verified".
+
+**Dependencies:** Phase 3.
+
+**Covers:** `remove-vault-sync.AC2.5`, `remove-vault-sync.AC2.6`, `remove-vault-sync.AC2.7`,
+`remove-vault-sync.AC5.1`–`remove-vault-sync.AC5.8`, `remove-vault-sync.AC6.4`,
+`remove-vault-sync.AC6.5`
+
+**Done when:** No user-visible string in `src/app/` or `src/components/` mentions the vault,
+syncing, or the bridge; AGENTS.md describes only code that exists; the simulator pass in AC6.4 and
+AC6.5 is complete with screenshots.
+<!-- END_PHASE_4 -->
+
+## Additional Considerations
+
+**Routine authoring narrows to one path.** With vault import gone, the AI Coach is the only way to
+create a routine, and it requires a user-supplied API key. There is no manual routine builder.
+This is an accepted consequence, not an oversight, and it is why the Today tab's empty state is
+rewritten rather than merely trimmed — that screen is now the sole onboarding funnel for a user
+with no routines.
+
+**Test coverage decreases.** Roughly 1,600 lines of test code are deleted. Three replacement tests
+are specified (deletable routine, v3→v4 migration, legacy settings blob) and the `upsertRoutine`
+defaulting coverage is verified rather than assumed, but net coverage still drops. That is the
+expected outcome of removing a feature and should not be read as regression.
+
+**Layout is invisible to the test suite.** AGENTS.md warns that `src/components` and `src/app` are
+covered by no jest project — a green run proves nothing about them. Phases 1 and 4 both edit
+screens, so each needs a simulator pass, not just a passing suite.
+
+**`../workout-bridge` is untouched.** The separate Node/vitest repo keeps its copy of the markdown
+contract and continues to work against any client that speaks it. Only this repo's obligation to
+mirror contract changes into it is removed. Archiving or deprecating that repo is out of scope.
+
+**Ordering within Phase 1 matters.** `onCompleteSession` runs the debrief hook last, after the
+session record closes and the HealthKit write is under way. Removing the sync block from the
+middle of that sequence must not reorder the surrounding effects.