diff --git a/.planning/research/v1427-r4-senior-dev.md b/.planning/research/v1427-r4-senior-dev.md
new file mode 100644
index 000000000..4c058773b
--- /dev/null
+++ b/.planning/research/v1427-r4-senior-dev.md
@@ -0,0 +1,619 @@
+# R4 senior-dev architectural review — v1.4.27
+
+Read-only review of the v1.4.27 MB1-MB7 surface on `develop` (HEAD
+`617d4518`). Scope frame from the dispatch directive; severity tiers
+follow `BLOCKER` → `HIGH` → `MED` → `LOW` → `NIT` with the v1.4.28
+backlog as the deferral target for non-blocking items.
+
+Conventions observed end-to-end: no forbidden vocabulary, Marc-voice
+English, no PII. The review prose follows the same conventions.
+
+## Summary
+
+The MB1-MB7 work is structurally sound. The new primitives compose with
+the existing shadcn vocabulary, the SSR-safe hooks degrade correctly,
+the migration is idempotent, and the context-based Coach launch model
+removes the sub-page unmount bug cleanly. The findings below are
+**zero blockers, two HIGH, four MED, six LOW**, plus one annotation
+that should land in the CHANGELOG.
+
+The single most consequential finding is **HIGH-1 (ResponsiveSheet
+footer slot under-used)** — every call site outside `export-section`
+inlines the form footer into the body instead of feeding it through the
+`footer` prop, so the sticky-pinned bottom rail the primitive promises
+on the Sheet branch only paints empty. The fix is a 5-minute swap per
+call site and lifts the actual UX promise to where the primitive's
+JSDoc and tests already claim it lives.
+
+---
+
+## BLOCKER
+
+(none)
+
+---
+
+## HIGH
+
+### HIGH-1 — `<ResponsiveSheet>`'s sticky footer rail is dead code on every consumer except `export-section`
+
+**Where:** every ResponsiveSheet call site:
+- `src/app/page.tsx:523, 533`
+- `src/app/measurements/page.tsx:92`
+- `src/app/mood/page.tsx:51`
+- `src/app/medications/page.tsx:261`
+- `src/components/medications/SideEffectsSection.tsx:327` (inline footer at line 462)
+- `src/components/medications/inventory-section.tsx:429` (inline footer at line 512)
+- `src/components/medications/intake-history-list.tsx:539, 687`
+
+**Symptom:** the primitive's docstring (`responsive-sheet.tsx:71-75`) and
+smoke suite both pin a sticky-pinned `<SheetFooter>` on the Sheet
+branch so primary CTAs stay reachable when the keyboard collapses the
+viewport. That's the entire point of the bottom-sheet bias on mobile.
+None of the call sites listed above pass a `footer={…}` prop. They
+inline the Save / Cancel row inside the `children` body. The body has
+`overflow-y-auto` (line 138 of the primitive), so on phones the CTA row
+scrolls **away** under the keyboard exactly like the pre-MB1 baseline.
+
+`grep -rn "footer=" src --include="*.tsx"` in this tree returns four
+hits, all in `src/components/settings/export-section.tsx` — and that
+file does not use `<ResponsiveSheet>` at all (the `footer` token there
+binds to a different component).
+
+**Fix:** the form-owner moves the footer JSX out of the form body and
+passes it through `footer={…}`. The `<form>` tag stays in the body
+because the `Save` button has `type="submit"` and form-association
+crosses sticky-pinning fine when the submit button shares the same form
+ancestor. Either move the `<form>` to wrap the whole `<ResponsiveSheet>`
+or use `form="<id>"` on the submit button. Five call sites; ~80 lines
+moved total.
+
+**Impact:** the WCAG 2.5.5 floor work + the rest of MB1's mobile
+narrative banks on this. Today the floor is correct but the rail is
+unreachable when the keyboard is up — same UX class the lift was meant
+to retire.
+
+**Tier:** HIGH because the primitive ships a contract its consumers
+silently violate. The fix is mechanical, not architectural; v1.4.28
+W1a candidate.
+
+### HIGH-2 — `<ResponsiveSheet>` mounts Sheet AND Dialog at the breakpoint boundary, losing focus + scroll position on viewport rotation
+
+**Where:** `src/components/ui/responsive-sheet.tsx:96, 158`
+
+**Symptom:** the primitive's render branch is a hard `if (isMobile)
+return <Sheet>…; return <Dialog>…;`. Both branches are full Radix
+portal mounts. When the viewport crosses 768 px (tablet rotation, browser
+window resize, devtools toggle) `useIsMobile` flips and React unmounts
+one tree and mounts the other.
+
+For an open ResponsiveSheet that means:
+- input focus is lost (the focused element is unmounted)
+- a partially-scrolled body resets to the top
+- a half-typed `<textarea>` value is preserved only because every
+  consumer holds the value in React state, not the DOM — but that's a
+  consumer-side rescue, not a primitive guarantee
+
+**Fix proposals (pick one for v1.4.28):**
+1. Lock the branch at first mount: `const [isMobile] = useState(() =>
+   useIsMobile())` so the chosen mount survives rotation. Trade-off:
+   a user who rotates while the sheet is open keeps the wrong layout
+   for that session. Acceptable for the dominant use case (rotation
+   while a sheet is open is rare).
+2. Match-route the same `<DialogPrimitive.Root>` against a different
+   `<Content side=…>` driven by Tailwind responsive classes, so the
+   tree never unmounts. Higher refactor cost; better UX.
+
+**Tier:** HIGH because the breakage class (focus loss, scroll reset) is
+visible on a documented user flow (iPad portrait → landscape with the
+quick-entry sheet open). Not a v1.4.27 blocker because the trigger is
+narrow; flag for v1.4.28.
+
+---
+
+## MED
+
+### MED-1 — Three remaining raw `<select>` blocks bypass `<NativeSelect>`
+
+**Where:** `src/components/settings/ai-section.tsx:370, 823, 1008`
+
+**Symptom:** MB7 / CF-52 consolidated three Settings native pickers
+behind the shared `<NativeSelect>` primitive
+(`src/components/ui/native-select.tsx`) to retire the drifting
+`NATIVE_SELECT_CLASS` triplet. Three more raw `<select className="…">`
+still live in `ai-section.tsx`. The classNames differ from the primitive
+(`mt-1 h-9 w-full rounded-md border px-3 text-sm sm:max-w-md` vs the
+primitive's `flex h-9 w-full rounded-md border px-3 py-1 text-sm shadow-xs
+transition-[color,box-shadow]…`). That's exactly the drift the
+primitive was introduced to cure.
+
+**Fix:** swap to `<NativeSelect>` and add `mt-1 sm:max-w-md` via
+`className`.
+
+**Tier:** MED — visual drift Marc has already declared a bug class
+in MB7; not a regression in this release but the primitive's purpose is
+defeated as long as these three lag.
+
+### MED-2 — `CoachLaunchScope` parameter is silently dropped today; no schema lock for v1.4.28
+
+**Where:** `src/lib/insights/coach-launch-context.tsx:33-73`
+
+**Symptom:** `askCoach(prefill, scope?)` accepts the scope arg and
+explicitly `void scope`s it, documenting that v1.4.28 will pre-narrow
+the sources rail. Good forward-compat shape, but:
+1. The `CoachLaunchScope` type is `{ metric?: string }` — `string` is
+   not narrowed to the metric-key union the sources rail consumes
+   (`CoachScopeSource`). When v1.4.28 wires it up the call sites can
+   already pass any string and only typecheck against `string`.
+2. No call site passes scope today (`grep "askCoach\(" src` returns 4
+   call sites; all single-argument). Zero usage, zero compile-time
+   pressure to keep the contract honest.
+3. The `scope?.metric` shape will not survive contact with a "narrow to
+   medication + mood" multi-source request, which the v1.4.28 plan
+   already hints at.
+
+**Fix for v1.4.27:** tighten `metric?: string` to the existing
+`CoachScopeSource` (or `CoachScopeSource[]`) type imported from
+`@/lib/ai/coach/scope` (or wherever it lives) so the future wiring is
+type-locked.
+
+**Tier:** MED — forward-compat hazard, no runtime cost today.
+
+### MED-3 — `<ResponsiveSheet>` Dialog branch ignores `bodyClassName` for the footer; visual drift between branches
+
+**Where:** `src/components/ui/responsive-sheet.tsx:179-191`
+
+**Symptom:** the Sheet branch's footer uses
+`sticky bottom-0 mt-0 flex-row justify-end gap-2 border-t bg-background/95
+backdrop-blur supports-[backdrop-filter]:bg-background/80`. The Dialog
+branch's footer is `<DialogFooter>` with no class override. They render
+visually different on a tablet portrait that just crossed 768 px going
+wider — a different border, a different background. The
+`data-slot="responsive-sheet-footer"` selector is identical on both
+branches so test assertions against the slot pass uniformly even though
+the visual contract diverges.
+
+**Fix:** either pull a shared `RESPONSIVE_SHEET_FOOTER_CLASS` constant
+out so both branches stay synchronized, or document the divergence in
+the primitive's JSDoc as intentional (the smoke suite already pins this
+shape at line 173-188 of the test file).
+
+**Tier:** MED — visual drift class; same primitive Marc just introduced
+to cure visual drift elsewhere.
+
+### MED-4 — Empty-state Insights sub-pages drop the Coach launch surface
+
+**Where:** every `/insights/{metric}/page.tsx` empty branch — confirmed
+in `blutdruck/page.tsx:77-92`, `gewicht/page.tsx:75-…`, etc.
+
+**Symptom:** when the user has zero observations for a metric, the
+sub-page short-circuits to a `<SubPageShell>` + `<EmptyState>` and
+**does not** mount the `<CoachLaunchButton>`. The Coach is reachable
+from the layout's drawer mount, but the sticky FAB / inline-button pair
+on the populated branch is the only visible entry point on these pages.
+The empty branch leaves an unauthenticated-feeling page with one CTA
+("add a measurement") and no path to the Coach who can explain *why*
+adding the measurement matters.
+
+The mother `/insights/page.tsx` empty branch (line 140-155) does the
+same thing — Coach-less.
+
+This is the directive's "fallback CTA" question: today there is no
+secondary fallback. The empty-state is a single-CTA dead-end.
+
+**Fix proposal (v1.4.28):** add `<CoachLaunchButton />` below the
+`<EmptyState>` action so a user with zero data still has one tap to
+"ask the coach why blood pressure matters at my age". The Coach drawer
+already handles zero-data conversations (it just acknowledges the lack
+of data and falls back to general guidance).
+
+**Tier:** MED — explicit directive question, not a regression but a
+deliberate one-CTA call. Worth promoting to v1.4.28 if Marc agrees the
+Coach is the secondary CTA on these pages.
+
+---
+
+## LOW
+
+### LOW-1 — `useIsMobile` SSR-default biases mobile users toward a desktop first paint
+
+**Where:** `src/hooks/use-is-mobile.ts:21`
+
+**Symptom:** the hook returns `false` on the server **and** the first
+client render (`useState<boolean>(false)`). The effect then flips to the
+live value on the next tick. For a mobile user this means:
+1. Server renders the Dialog branch (centred, sm:max-w-md).
+2. Client hydrates that markup.
+3. Effect runs, hook returns `true`, Dialog branch unmounts, Sheet
+   branch mounts.
+
+Mobile users see one frame of the desktop variant before it flips. The
+visible flash is short (the effect runs synchronously after hydration)
+but it's there, and on a cold cache it's noticeable enough to matter
+for the "feels native" bar Marc set for v1.4.27.
+
+**Fix:** either accept the trade-off (the hook's doc-comment explicitly
+calls it out as the SSR-vs-flash trade), or read `navigator.userAgent`
+on first paint in a `useSyncExternalStore` wrapper so the initial value
+matches the live viewport. The `useSyncExternalStore` pattern keeps SSR
+markup deterministic while the client's first paint reads the live
+state.
+
+**Tier:** LOW — known trade-off; not a regression; the SSR-safe path
+is the right default for HealthLog's auth-gated SPA where most surfaces
+mount inside `"use client"` islands.
+
+### LOW-2 — `Checkbox`'s `onCheckedChange` boolean type bleeds the Radix `boolean | "indeterminate"` union to consumers
+
+**Where:** `src/components/ui/checkbox.tsx`, consumed at
+`src/components/insights/coach-panel/sources-rail.tsx:249`
+
+**Symptom:** the consumer calls `onCheckedChange={() => toggleSource(row.key)}`
+ignoring the value. The Radix `onCheckedChange` signature is
+`(checked: boolean | "indeterminate") => void`. The checkbox can never
+actually emit `"indeterminate"` here (no state pushes it there), so the
+consumer pattern is safe, but the type-narrowing is asymmetric: if a
+future consumer reads `checked` and forwards to `setState<boolean>` they
+need to narrow.
+
+**Fix:** consumers in v1.4.27 are clean; for v1.4.28 (or whenever a
+non-toggling checkbox lands) wrap the primitive's `onCheckedChange` to
+re-emit a strict `boolean` (treating `"indeterminate"` as `false`).
+
+**Tier:** LOW — no runtime issue; latent TS sharp edge.
+
+### LOW-3 — `Popover` lacks an `<Anchor>` re-export
+
+**Where:** `src/components/ui/popover.tsx:55`
+
+**Symptom:** exports `Popover, PopoverTrigger, PopoverContent`. The
+Radix `Popover` namespace also surfaces `Anchor`, useful when the
+trigger is icon-only and the popover should align to a different
+element (a label, a parent row). The Coach composer hint (the MB3
+consumer) doesn't need it today; the next consumer will.
+
+**Fix:** re-export `PopoverAnchor` from the primitive for parity with
+the rest of the shadcn surface.
+
+**Tier:** LOW — additive nit, not a defect.
+
+### LOW-4 — Visual-viewport listener does not debounce; rapid `auto` scrolls during keyboard animation
+
+**Where:** `src/components/insights/coach-panel/message-thread.tsx:208-221`
+
+**Symptom:** the iOS soft-keyboard animation fires several `resize`
+events as the keyboard slides up (each ~16 ms). Each fires
+`scrollTo({ behavior: "auto" })` which forces a sync scroll. The result
+is correct (the bubble lands at the tail) but the cost is several layout
+reflows during the animation. RAF-debouncing would coalesce them to one
+write per frame.
+
+**Fix:** wrap the resize handler in a `requestAnimationFrame` so
+multiple resize events per frame collapse to one scroll. Cleanup also
+needs to `cancelAnimationFrame(handle)`.
+
+**Tier:** LOW — perf nit, not a correctness bug. The cleanup itself is
+correct (lines 219-220).
+
+### LOW-5 — `ResponsiveSheet` swallows `className` collisions on the Sheet branch (`flex…` defaults vs consumer `className`)
+
+**Where:** `src/components/ui/responsive-sheet.tsx:104-111`
+
+**Symptom:** Sheet branch's container has a hardcoded `flex max-h-[90dvh]
+flex-col gap-0 rounded-t-2xl p-0` followed by `className`. The class
+order means consumer overrides for `max-h`, `rounded-t-*`, or `flex-*`
+have to fight Tailwind's specificity. The Dialog branch only adds
+`sm:max-w-md` so it composes cleanly. Marc's `medications/page.tsx:269`
+already passes `className="sm:max-w-lg"` — that lands on both branches
+because `sm:max-w-*` doesn't conflict with the Sheet's `max-h-[90dvh]`,
+but a future consumer who wants `max-h-screen` on the Sheet branch
+will need `!important`.
+
+**Fix:** route the consumer `className` through `cn(…)` with the
+defaults *after* it, or split the Sheet defaults into a separate
+`sheetClassName` prop. The smoke suite would catch the regression as
+soon as a consumer tests it.
+
+**Tier:** LOW — latent override friction, no consumer hits it today.
+
+### LOW-6 — `CoachLaunchProvider`'s prefill drop on close races a `react-strict-mode` double-invoke
+
+**Where:** `src/lib/insights/coach-launch-context.tsx:75-81`
+
+**Symptom:** `handleSetOpen(false)` calls `setPrefill(null)` immediately.
+Under React 19 + strict mode double-invoke a third-party listener that
+fires `setOpen(false)` twice would set `prefill` to `null` twice — a
+no-op, but it also tears down the cached prefill before the drawer's
+exit animation finishes reading it. The drawer renders the prefill via
+the `prefill` prop, and `CoachDrawer` already uses `useResettableValue`
+(line 168) so an immediate clear of `prefill` to `null` would seed a
+fresh empty input mid-close-animation if the close happens to interrupt
+a re-open with a prefill.
+
+**Fix:** defer the prefill clear to the next tick (`setTimeout(() =>
+setPrefill(null), 0)`) or wait for the drawer's transition-end before
+clearing. The race is narrow today (no observed bug) but the right place
+to fix it is in the provider, not the consumer.
+
+**Tier:** LOW — speculative; documents the trade-off.
+
+---
+
+## NIT
+
+### NIT-1 — `ResponsiveSheet`'s test mock-tree includes both portal mocks even though the test exercises one branch
+
+**Where:** `src/components/ui/__tests__/responsive-sheet.test.tsx:20-116`
+
+The mock setup runs at module-evaluate time (`vi.mock` calls hoist) so
+both Sheet and Dialog mocks are installed regardless of which branch
+the assertion exercises. It works fine, but trims to roughly half its
+size if each `describe.each([…])` block sets up only the branch it
+tests. Future readers will appreciate the smaller mock surface.
+
+### NIT-2 — `useIsMobile` defaults to the `"md"` token; the Coach drawer's `useIsMobile("sm")` call is the only `"sm"` consumer
+
+**Where:** `src/components/insights/coach-panel/coach-drawer.tsx:142`
+
+The hook's `breakpoint: "sm" | "md"` parameter is a binary today. If a
+future consumer wants `lg` or `xl` the union has to expand. Documenting
+this as a known limitation (or generalising to a numeric `px`
+parameter) saves a future patch round.
+
+---
+
+## Migration safety — `0061_audit_log_carrier`
+
+The migration is **safe to deploy**. Verified against the project's
+established pattern:
+- Same `IF NOT EXISTS` shape as `0058_user_research_mode` (the
+  migration's own docstring cites the parallel).
+- Both columns are nullable so older rows, private-IP rows, and
+  offline-miss rows stay valid (no backfill needed).
+- The schema row at `prisma/schema.prisma:1240-1241` matches the SQL
+  exactly (`asn Int?` ↔ `INTEGER`, `carrier String?` ↔ `TEXT`).
+- Prisma client regeneration: the Dockerfile builder stage runs
+  `pnpm db:generate` at line 18, and the generated tree in
+  `src/generated/prisma/internal/prismaNamespace.ts` already lists
+  `auditLog` in the model props (`AuditLogOmit` shape line 4526). The
+  runtime client at `docker-entrypoint.sh:47-48` runs `prisma migrate
+  deploy` against the live DB before the app starts. Sequence is
+  airtight.
+- `src/lib/auth/audit.ts:48-50` reads `asnRow.asn` / `asnRow.carrier`
+  and writes to the new columns inside a fire-and-forget race wrapped
+  in a 3 s timeout (line 28-38). Failure modes:
+  - Online provider hung → 3 s timeout → location backfill skipped, ok.
+  - MMDB miss → `lookupIpAsn` returns null → audit row stays without
+    asn/carrier, ok.
+  - Race shape: `Promise.race([Promise.all([…]), timeout])` — the
+    location lookup happens inside the `Promise.all` arm. If the
+    location lookup hangs, the timeout arm wins and the `.then(result
+    => …)` short-circuits on `if (!result) return;`. ASN lookup is
+    synchronous (`Promise.resolve(lookupIpAsn(...))`) so it never
+    contributes to the race delay. Sound.
+
+The one **observation** (not a finding): the `auditLog()` helper
+returns the original `prisma.auditLog.create` result before the
+fire-and-forget runs. Callers that `await` `auditLog()` and then
+immediately read `auditLog.findUnique(…)` will see a row without
+`asn`/`carrier`. No call site does this today, but tests that mock
+both `lookupIpLocation` and `lookupIpAsn` need to flush the microtask
+queue before asserting (the audit test suite at line 191-… already
+does so).
+
+---
+
+## Drift between `<ResponsiveSheet>` and existing `<Sheet>` / `<Dialog>`
+
+Sheet surface still has these direct mounts:
+- `src/components/insights/coach-panel/coach-drawer.tsx:265` — raw
+  `<Sheet>` driven by `useIsMobile("sm")` that flips `side` between
+  `"bottom"` and `"right"` (not Sheet vs Dialog). Distinct pattern
+  from MB1's "Sheet on mobile, Dialog on desktop" — both ship today
+  side-by-side.
+- `src/components/insights/coach-panel/coach-drawer.tsx:493, 525` — the
+  history + sources rail trays, `side="left"` / `side="right"`. These
+  are *adjacency* surfaces, not primary modals; correct to stay on raw
+  Sheet.
+- `src/components/insights/coach-panel/coach-settings-sheet.tsx` — same
+  adjacency pattern.
+- `src/components/layout/bottom-nav.tsx` — the mobile nav drawer; not a
+  form modal.
+
+Dialog surface has 11 remaining direct consumers
+(`grep -l "from \"@/components/ui/dialog\""`). The notable ones are:
+- `phase-config-dialog.tsx` — could migrate to ResponsiveSheet (form on
+  mobile would benefit). v1.4.28 candidate.
+- `ResearchModeAcknowledgmentDialog.tsx` — modal disclaimer with primary
+  CTA; benefits from the bottom-sheet on phones.
+- `target-edit-sheet.tsx` — already imports Dialog; the file name lies.
+  Consolidation target.
+- `mood-list.tsx`, `measurement-list.tsx` — edit-row modals. The mood
+  and measurement *create* surfaces flipped to ResponsiveSheet via
+  `<MoodForm>` / `<MeasurementForm>`, but the row-edit modals
+  did not. Asymmetric.
+
+**Net:** the drift is real and the MB1 work is incomplete in the
+direction the directive worried about. Not a v1.4.27 blocker because
+every remaining Dialog still works on mobile (it just keeps the
+pre-MB1 layout). The five Dialog → ResponsiveSheet swaps above are a
+natural v1.4.28 follow-up bucket.
+
+---
+
+## Form integrations to `<ResponsiveSheet>` — react-hook-form intact?
+
+The codebase is `react-hook-form` *installed* but not *used* in the
+form components that landed under MB1:
+- `MedicationForm`, `MeasurementForm`, `MoodForm`,
+  `SideEffectsSection`, `InventorySection`, `IntakeHistoryList` — all
+  `useState`-based. `grep -rln "useForm\b" src` returns zero hits.
+- `package.json:71` lists `react-hook-form: ^7.75.0` and
+  `@hookform/resolvers: ^5.2.2` as deps; nothing imports them.
+
+The form integration question is therefore vacuous — there are no
+react-hook-form forms to break. The `useState` pattern composes
+cleanly with ResponsiveSheet because `ResponsiveSheet` is
+state-passthrough only.
+
+(Tangent: the unused `react-hook-form` deps are a v1.4.28 hygiene item
+— either commit to one form library or drop the deps. Out of scope for
+this review.)
+
+---
+
+## `<Input>` `inputMode` derivation — controlled / uncontrolled safety
+
+`src/components/ui/input.tsx:50-78` adds a single derived attribute
+(`inputMode`) and does not touch `value` / `defaultValue` / `onChange`.
+Both controlled and uncontrolled `<input>` callers pass through
+unchanged. The smoke suite at `src/components/ui/__tests__/input.test.tsx`
+covers:
+- Default derivation for every type the switch handles.
+- Explicit `inputMode="numeric"` override on `type="number"` (line 64-70).
+- Plain text drops `inputMode` entirely (line 59-62).
+
+One thin edge: the `type` prop is `React.HTMLInputTypeAttribute |
+undefined`. If a future consumer passes `type={undefined}` and
+`inputMode={undefined}` the primitive emits `inputMode={undefined}`
+(omitted attribute), which is the React-clean shape. No risk.
+
+The pattern correctly leaves password-manager hints intact for explicit
+`autoComplete` values (line 57-58). Already covered by the existing
+test.
+
+---
+
+## `LayoutCoachMount` boundary
+
+`src/components/insights/layout-coach-mount.tsx` is 26 lines, one
+`useCoachLaunch()` hook call, one `<CoachDrawer>` mount. The boundary
+between the server-component layout (`src/app/insights/layout.tsx`)
+and the client-only `<CoachDrawer>` is correctly threaded:
+- `layout.tsx` stays server-renderable (no `"use client"` directive,
+  but the children + the provider + the mount are all marked
+  `"use client"`).
+- `LayoutCoachMount` is the bridge — tiny client island that consumes
+  the context.
+- `<CoachLaunchProvider>` ships its own `"use client"` directive
+  (`coach-launch-context.tsx:1`).
+
+The boundary is at the right level. The only nit is that the bridge
+component could trivially inline into `CoachLaunchProvider` (the
+provider could render the drawer alongside `{children}`), but that
+couples the provider to the drawer mount and would block the v1.4.28
+case where a non-Coach client consumes the same context (e.g. a future
+"Ask the coach" CTA on `/dashboard`). The current shape stays open for
+that.
+
+---
+
+## `<CoachLaunchProvider>` + `useCoachLaunch` — React 18 patterns review
+
+The context shape is textbook for React 19:
+- `useState` for `open` + `prefill` (each independent state cell).
+- `useCallback` on `askCoach` + `handleSetOpen` so consumers don't
+  re-mount on every parent render.
+- `useMemo` on the `value` so the context never invalidates unless a
+  state cell flips.
+- Reading the context with `useCoachLaunch()` returns `null` outside
+  the provider so consumer fallbacks (the launch button's null-render,
+  the mother page's conditional `onAskCoach`) compose cleanly.
+
+Stale-closure traps: none in the current code path. `askCoach`'s body
+references `setPrefill` and `setOpen` only (setters are stable). The
+`scope` arg is currently `void`ed. If v1.4.28 starts using `scope`
+inside `askCoach` to update sources-rail state, that state must be
+lifted into the provider or read via a setter-pattern hook — not closed
+over.
+
+The `useMemo` deps array correctly lists every state cell + every
+callback. No drift.
+
+---
+
+## Commit attribution drift in MB1 — release blocker or annotation?
+
+The drift is real and limited to **one** commit:
+
+**`bd7cb938` carries the wrong subject + body.** The commit message
+reads `feat(ui): lift the password toggle into a shared primitive`
+with a `<PasswordInput>` move narrative. The diff is the
+ResponsiveSheet migration of `src/app/page.tsx`, `src/app/measurements/page.tsx`,
+`src/app/mood/page.tsx` (Dashboard / Measurements / Mood pages swapping
+`<Dialog>` for `<ResponsiveSheet>` on the quick-entry sheets). The
+*real* PasswordInput move landed in `b25683404` six minutes later with
+the identical commit body. So the body was duplicated, the diff is the
+wrong work.
+
+**`44554729` is fine.** The commit message
+`chore(insights): apply the tap-target floor across the insights surface`
+matches its diff (coach-drawer.tsx pills + buttons lifted to `h-11`,
+hero-strip drops `size="sm"`, etc.). The directive's flag here looks
+like a misread — this commit's narrative is internally consistent.
+
+**Release-blocker assessment:** not a blocker. The commit hash, author,
+date, and merge tree are intact; the only damage is to the git log's
+storytelling. The fix options:
+1. Annotate v1.4.27 CHANGELOG with a note: "commit bd7cb938 ships the
+   ResponsiveSheet migration of Dashboard / Measurements / Mood quick-
+   entry sheets under a misattributed subject — see b25683404 for the
+   intended PasswordInput primitive move."
+2. Cherry-pick the right body onto a fresh commit with `git commit
+   --amend` (rewrites history; develop branch only — not yet on main
+   per the branch model). The MB1 PR is already on develop so an amend
+   forces a `git push --force-with-lease`. Per the safety protocol Marc
+   would have to authorise.
+3. Leave it; the diff is reviewable, the tag is months away.
+
+**Recommendation:** CHANGELOG annotation. v1.4.27 ships from develop →
+main with the misattributed commit intact + the annotation explaining
+the swap. No history rewrite. Marc's PII-aware release notes already
+gloss over commit subjects; the audit trail stays in the planning
+docs.
+
+---
+
+## Cross-cutting observations (not severity-tiered)
+
+- **Test coverage is healthy.** ResponsiveSheet smoke at 6 assertions,
+  coach-launch-context smoke at 3 assertions, input.tsx at 9
+  assertions. The patterns + scenarios that matter are pinned.
+- **i18n surface is clean.** No new untranslated keys leaked into the
+  primitives (they're all consumer-driven `t(...)` calls).
+- **No PII / personal data in any of the new strings.** The migration
+  SQL comments reference example carriers ("Deutsche Telekom AG") but
+  the columns themselves are arbitrary user-traffic data, not the
+  reviewer's own.
+- **The forbidden vocabulary scan is clean.** `grep -rn -E
+  "AI|Claude|agent|marathon|wave|phase|session|subagent|Anthropic" src`
+  against the v1.4.27 diff returns only `phase` matches inside data
+  variables (`MedicationPhase`, `phase-config-dialog`, etc.) — domain
+  vocabulary, not authorship references.
+
+---
+
+## Suggested follow-ups (v1.4.28 candidates)
+
+1. **HIGH-1 fix** — move inline form footers to `<ResponsiveSheet
+   footer={…}>` across all five call sites. ~80 LOC.
+2. **HIGH-2 path** — pick rotation strategy for ResponsiveSheet
+   (lock-at-mount vs unified Dialog/Sheet root).
+3. **MED-1 fix** — swap `ai-section.tsx`'s three raw `<select>` blocks
+   to `<NativeSelect>`.
+4. **MED-2 fix** — type-narrow `CoachLaunchScope.metric` to the
+   `CoachScopeSource` union before v1.4.28 wires it up.
+5. **MED-3 fix** — extract `RESPONSIVE_SHEET_FOOTER_CLASS` constant or
+   document the divergence as intentional.
+6. **MED-4 decision** — secondary CTA on Insights empty surfaces; if
+   yes, mount `<CoachLaunchButton>` below the action.
+7. **Drift cleanup** — convert 5 remaining `<Dialog>` consumers
+   (`phase-config-dialog`, `ResearchModeAcknowledgmentDialog`,
+   `mood-list` row-edit, `measurement-list` row-edit, `target-edit-sheet`)
+   to `<ResponsiveSheet>`.
+8. **react-hook-form deps** — either commit to the lib or drop the two
+   unused deps.
+
+---
+
+End of review.