diff --git a/.planning/STATE.md b/.planning/STATE.md
index c6b1ab4ef..304bb12a6 100644
--- a/.planning/STATE.md
+++ b/.planning/STATE.md
@@ -1,6 +1,6 @@
 # v1.4.22 marathon — state log
 
-Status: Wave 2 — Insights surface polish complete
+Status: Wave 3 — Coach polish complete
 Last update: 2026-05-10T21:00+02:00
 
 > Previous patch: v1.4.21 live (image digest
@@ -82,36 +82,38 @@ rewrite needs research first.
 
 ### B1 — Prompt rewrite
 
-- [ ] Adopt W1b research output: natural prose-first, motivational,
+- [x] Adopt W1b research output: natural prose-first, motivational,
       conservative, evidence-grounded
-- [ ] PROMPT_VERSION 4.22.0; EN + DE
-- [ ] Coach-specific test suite covering tone + evidence-collapsible
+- [x] PROMPT_VERSION 4.22.0; EN + DE
+- [x] Coach-specific test suite covering tone + evidence-collapsible
       pattern
+- Detailed report: `.planning/phase-W3-report.md`
 
 ### B2 — Collapsible provenance + values
 
-- [ ] Move inline raw values out of prose into a
+- [x] Move inline raw values out of prose into a
       "Worauf bezieht sich das?" expandable below each assistant
-      message
-- [ ] Source chips stay visible; values reveal on click
+      message; `---KEYVALUES---` … `---END---` sentinel stripped
+      server-side, parsed into `provenance.keyValues`
+- [x] Source chips stay visible; values reveal on click
 
 ### B3 — User avatar = Gravatar (parity)
 
-- [ ] Same dimensions as Coach avatar
-- [ ] Reuses existing `gravatarUrl` from `/api/auth/me`
+- [x] Same dimensions as Coach avatar
+- [x] Reuses existing `gravatarUrl` from `/api/auth/me`; initials
+      fallback when no Gravatar configured
 
 ### B4 — Disclaimer move
 
-- [ ] "Coach replies are generated. Clinical decisions belong with
-      your doctor." moves from below input → right rail
-- [ ] Right rail "Coach reads only your connected data" repurposed
-      so the source picker stands alone
+- [x] "Coach replies are generated. Clinical decisions belong with
+      your doctor." moves from below input → sources rail footer
+- [x] "Coach reads only your connected data" caption replaced by the
+      disclaimer so the source picker stands alone
 
 ### B5 — Settings cog wire-or-remove
 
-- [ ] Decide: wire to per-user prompt-tuning toggles (Coach tone,
-      verbosity), OR remove the icon
-- [ ] No dead buttons in v1.4.22
+- [x] Removed for v1.4.22; per-user prompt-tuning surface deferred
+      to v1.4.23 (no dead buttons in this release)
 
 ## Wave 4 — Other surfaces + backlog cleanup
 
diff --git a/.planning/phase-W3-report.md b/.planning/phase-W3-report.md
new file mode 100644
index 000000000..b3ab303e7
--- /dev/null
+++ b/.planning/phase-W3-report.md
@@ -0,0 +1,114 @@
+# Wave 3 — Coach polish report (v1.4.22)
+
+Adopts the W1b research output: the Coach now talks like a partner
+who sits alongside the user, not a database cursor. Load-bearing
+numbers move out of the prose into a collapsible block at the bottom
+of every assistant turn. Five atomic commits on `origin/develop`.
+
+## Commits
+
+1. `feat(coach): warm motivational rewrite with evidence-block sentinel`
+   — system-prompt EN + DE rewrite, PROMPT_VERSION 4.20.2 → 4.22.0,
+   broadens schema-test regex off the 4.20.x pin.
+2. `feat(coach): parse evidence-block sentinel out of streamed prose`
+   — `---KEYVALUES---` … `---END---` parser, `CoachProvenance.keyValues`
+   extension, route + persistence + SSE provenance frame wiring,
+   integration coverage.
+3. `feat(coach): collapsible evidence-block disclosure under assistant
+   bubbles` — `<details>` disclosure with chevron in
+   `message-thread.tsx`; closed by default; hides when keyValues is
+   empty; both EN + DE labels.
+4. `refactor(coach): drawer polish — Gravatar avatar, disclaimer
+   move, cog removal` — user bubble pulls `useAuth().user.gravatarUrl`
+   at the same 32-px dimensions as the Coach avatar with initials
+   fallback; disclaimer moved out of `coach-input.tsx` into
+   `sources-rail.tsx`'s footer; settings cog removed cleanly from the
+   drawer header.
+5. `chore(planning): record Wave 3 Coach polish completion` — STATE
+   tick + this report.
+
+## Sentinel-parse design
+
+Wave 3 ships a text-marker sentinel (`---KEYVALUES---` … `---END---`)
+with line-formatted entries (`<label>: <value> [<unit>] (<window>)`)
+rather than the fenced-JSON shape W1b sketched. The maintainer
+called this directly — robust against accidental Markdown-render
+collisions, and easier to scan in raw provider logs.
+
+Hard caps stack three ways:
+
+- 1 KB on the sentinel-block payload (post-`---KEYVALUES---` byte
+  count). Prompt-injection padding cannot grow the persisted
+  envelope.
+- 8 lines kept after parsing — matches the prompt contract.
+- Each line validates against the Zod `coachKeyValueSchema`
+  (length-capped label / value / unit / window). Malformed lines
+  drop silently; the route logs a wide-event
+  `coach.keyvalues.parse_failed` only when the block as a whole is
+  malformed (missing close sentinel, truncated, or zero valid rows).
+
+When the provider returns prose-only (mock client, refusal path,
+legacy provider that hasn't been re-prompted), the parser is a
+no-op: prose passes through, no `keyValues` field lands on the
+persisted envelope, and the UI hides the disclosure entirely.
+
+## Settings cog
+
+Removed per the maintainer's directive. The placeholder tooltip
+("Coach settings arrive in v1.4.21") read as a dead button after
+v1.4.21 shipped without delivering that surface. The v1.4.23
+candidate is a per-user prompt-tuning panel (tone, verbosity,
+fallback-language preference) — when that's ready, the cog returns
+with real wiring.
+
+## Concrete prose example
+
+User asks: "Meinst du, ich sollte mehr Sport machen?"
+
+The new prompt produces (synthesised from the prompt's rules + the
+embedded few-shot):
+
+> Bewegung sehe ich in deinem Tracking gerade nicht — magst du mir
+> kurz erzählen, wie eine typische Woche bei dir aussieht?
+> Spaziergänge, Sport, irgendwas Strukturierteres? Dann schauen wir
+> gemeinsam mit dem, was ich sehe, ob mehr Sport gerade Sinn ergibt.
+
+No sentinel block (the turn is qualitative, no numbers carry the
+answer), so the disclosure is hidden client-side. The persona
+("warm, neugierig, zurückhaltend"), the rule-3 invitation pivot, and
+the rule-5 open question all fire in three sentences.
+
+## Test count delta
+
+- Unit tests: 2068 → 2085 (+17 — 14 system-prompt assertions, 18
+  sentinel-parse cases, 4 message-thread disclosure + Gravatar
+  cases; offset against the disclaimer-test rewrite in coach-input
+  and sources-rail).
+- Integration tests: 5 → 7 (+2 — sentinel round-trip + prose-only
+  graceful-degrade against the postgres testcontainer).
+
+## Verification gates
+
+- `pnpm typecheck` — clean.
+- `pnpm lint` — 13 pre-existing warnings (no new entries).
+- `pnpm test --run` — 241 files / 2085 tests green.
+- `pnpm test:integration tests/integration/coach-chat.test.ts` —
+  7 tests green.
+
+## What did not ship
+
+Nothing was punted from the Wave 3 scope. The B3 Gravatar parity
+uses initials when the user's email is missing — the lookup is
+already server-side in `/api/auth/me`, so there's no extra round
+trip on the client. The Coach drawer mobile-tray test was untouched
+and still passes after the cog removal.
+
+## What stays as Wave 3 followups
+
+- v1.4.23 candidate: per-user Coach settings surface (tone slider,
+  language preference). Brings the cog back with real wiring.
+- The `mock-client.ts` provider remains prose-only by design — no
+  changes needed.
+- Sentinel format is documented in the prompt body itself; a
+  future docs-site page on the v1.4.22 release covers the
+  user-facing disclosure.