HealthLog/.planning/phase-W1a-report.md

Wave 1a — PROD probe + token-leak hunt + backlog inventory (v1.4.22)

Started: 2026-05-10T20:20+02:00 Branch: develop (post-v1.4.21) Production probed: https://healthlog.bombeck.io (image digest sha256:4e818d44702c…, v1.4.21) Probe scripts (read-only, no commits):

• .planning/v1422-w1a-tokens-probe.mjs (admin/api-tokens overflow walk)
• .planning/v1422-w1a-insights-probe.mjs (Insights Metrik / token leak hunt)
• .planning/v1422-w1a-targets-probe.mjs (Zielwerte page screenshot)

1. BD-Zielbereich tile

What I see today

src/components/charts/trend-card.tsx paints the BD tile in the shared <TrendCard> chrome:

• Headline (3xl bold): latest = data.bpInTargetPct — routed through windows.allTime?.pct since v1.4.19 A1. For the live tenant this is ~10.8 % rounded to 11 %.
• Sub-row (mt-auto muted text-xs): 7d: <avg7> · 30d: <avg30> — data.bpInTargetPct7d / pct30d. For the live tenant, both 50 %.
• No window label on the headline. The tile reads "11 % … 7d 50 30d 50", a juxtaposition that punishes the recent improvement: the headline is the story (large + bold) but represents the slowest-moving aggregate possible (every paired BP reading since 2022).

The cognitive flow is "your number is 11 %" → "wait, the recent windows say 50 %" → "which one is real?". Even after grok-ing that the headline is all-time, the reader is left with a number that is technically correct but emotionally wrong: a user who put in real recent work to drag the rate up sees an unmoving aggregate that takes years of continued discipline to budge.

Three proposals (recommend one)

A) Re-anchor headline to last-30d, surface all-time as a sub-value

Pro:

• The dashboard tile becomes a "what's happening now" widget, matching every other tile on the page (weight, pulse, mood — all show recent reading + 7d/30d). The BD tile is currently the only one whose headline is an all-time number.
• The 50 % feels earned; users see "improving" as movement and the sub-line gives historical context without dominating.
• One-line API change: bpInTargetPct = windows.last30Days?.pct plus a new field bpInTargetPctAllTime for the sub-row.

Con:

• Inverts the v1.4.19 A1 fix story. Need a backlog entry to clarify that the v1.4.19 fix was about correctness (the headline was silently equal to the 30-day) and v1.4.22 is about framing (the all-time aggregate belongs in the small text).
• Cache invalidation on the analytics query (existing TanStack pattern handles this).

B) Explicit window label "Allzeit / All time" on the headline

Pro:

• Smallest diff. Add a <span className="text-xs text-muted">{t( "dashboard.bpInTargetAllTimeLabel")}</span> next to the 11 %. No data-shape change, no API change.
• Honours the v1.4.19 A1 contract verbatim — the headline still surfaces the slowest-moving number.

Con:

• The cognitive dissonance remains: the user still reads the big number first and the label second. A 50 % → 50 % → 11 % "punishment reading" persists; the label only confirms why.
• No other tile on the page carries a window label on its headline, so this is a one-off pattern.

C) Three-window chip row (no headline) — replace big-number layout

Pro:

• Renders the tile as three equal-weight chips: 7d 50 % · 30d 50 % · Allzeit 11 %. No single number dominates; the user picks the window that matters to them.
• Visually distinct enough to break the "every TrendCard is identical" uniformity, which signals to the user that BP-in-target is a conceptually different metric than weight / pulse.

Con:

• Largest design lift. Diverges from the shared <TrendCard> chrome used everywhere else on the dashboard, so either (a) BD tile gets a bespoke <BpInTargetTile> component, or (b) the shared TrendCard picks up an optional "windowed mode". Either is a Wave-2/3 redesign item, not a Wave-4 polish patch.
• Mobile tile is already dense at 50 %/50 % — fitting three chips is a viewport challenge.

Recommendation: A (re-anchor to last-30d, all-time → sub-row)

Smallest UX win for the smallest engineering lift, and aligns the BD tile with every other tile on the dashboard. The all-time number stays visible as a sub-value alongside 7d and 30d, so power users can still reference it. The v1.4.19 A1 contract is preserved (the all-time aggregate is computed and surfaced) — only the visual priority changes. The fix is roughly:

• src/app/api/analytics/route.ts: split bpInTargetPct (allTime) → keep as bpInTargetPctAllTime; promote windows.last30Days.pct to the new bpInTargetPct headline.
• src/components/charts/trend-card.tsx — extend <TrendCard> with an optional tertiaryAvg slot OR (cleaner) accept a third sub-value pair (avgAllTime, avgAllTimeLabel) and render it as · Allzeit: <pct> % next to 7T: and 30T:. Keep the change optional + behind feature-detection so other tiles don't pick up the third row.
• Tests: bp-in-target.test.ts already pins the three-window contract; add a <TrendCard> test that asserts the headline matches the 30-day pct when both are passed.

2. Admin / api-tokens scrollbar

Per-viewport probe results

Viewport	viewport×inner	doc overflow	body overflow	Real culprit
iphone-se 375×667	375 / 375	NO	NO	(none — only the intentional nav.no-scrollbar mobile section strip)
pixel5 393×851	393 / 393	NO	NO	(none — same intentional nav)
ipad-mini 768×1024	768 / 768	NO	NO	<div class="hidden overflow-x-auto md:block"> (intentional auto, BUT triggers because the inner table overflows) — table sw=212, cw=122, delta=90
desktop 1280×800	1280 / 1280	NO	NO	Same wrapper. Table sw=663, cw=634, delta=29

Real culprit

The desktop table overflows its overflow-x-auto wrapper at md+ even on a 1280-wide desktop viewport.

Selector: div.hidden.overflow-x-auto.md:block → table.w-full.table-fixed.text-sm (file: src/components/admin/api-token-overview-section.tsx:145-238).

Why the table overflows even with table-fixed + percentage colgroup:

1. The "Last used" column (<col className="w-[12%]" />) carries whitespace-nowrap (line 225) and renders strings like "05.05.2026, 21:46" (~14 chars after locale) → ~110 px at the table's font size.
2. The "Created" column (<col className="w-[8%]" />) also carries whitespace-nowrap (line 230) and renders the same locale-aware "05.05.2026" ~80 px wide.
3. At 1280 px viewport with sidebar 220 px + gap 24 px + page padding ~64 px, the card's content area ≈ 700 px. 12 % of 700 = 84 px, 8 % = 56 px. The whitespace-nowrap strings (~110 + ~80 px) overflow their colgroup-allotted percentages.
4. whitespace-nowrap wins over table-fixed's width contract, so the table's intrinsic width exceeds 100 %. The wrapper's overflow-x-auto then paints a horizontal scrollbar to let the user see the clipped columns — that is the scrollbar Marc keeps reporting.

Visual confirmation: /tmp/v1422-w1a/desktop-fullpage.png shows the "Last used" and "Create" headers truncated mid-word and the rightmost date strings clipped (e.g. "05/05/" with the year cut off).

Concrete fix path

Three options, in order of preference:

1. Drop whitespace-nowrap on the date cells (lines 225, 230). formatDateTime("05.05.2026, 21:46") is two semantic chunks; let it wrap to two lines on narrow viewports. The card grows by one row but no scrollbar — and the desktop case at >1024 px already fits the unwrapped string in 12 %.

2. Compact date format on the "Last used" column. Use formatDateShort() (already in src/lib/format.ts:51) which drops the year. "05.05" + time fits in 70 px; the trailing tooltip already exposes the full timestamp on hover.

3. Hide the "Created" column on md (768-1023 px). Add a hidden lg:table-cell to the 6th <col> + <th> + <td>. The tablet user trades one column for a clean layout; desktop keeps everything.

Option 1 is smallest diff (two-line change, no i18n surface impact). Option 2 is the cleanest UX. Option 3 is what the "User"/"Permissions" columns already do via the mobile card-list — same pattern, one breakpoint up.

3. Raw metric:<TYPE> token leaks

Where they emit

The Playwright probe captured the literal substrings rendered in the production /insights page (German locale, live-tenant data):

Match	Context (rendered prose, last 60 + 60 chars)
metric:BLOOD_PRESSURE_SYS	"…strukturiertes Gegensteuern sinnvoll ist. metric:BLOOD_PRESSURE_SYS\n\nZuletzt aktualisiert"
metric:BLOOD_PRESSURE_SYS	"…Medikation mit niedrigeren systolischen Werten einhergeht. metric:BLOOD_PRESSURE_SYS\n\n…"
metric:WEIGHT	"…damit ich die BMI-Klassifikation präzise berechnen kann. metric:WEIGHT\n\nZuletzt aktualisiert"
metric:PULSE	"…und aktuell, daher ist die Aussage belastbar. metric:PULSE\n\nZuletzt aktualisiert"
metric:MOOD	"…nur 4 % Zielbereichstreffern in den letzten 30 Tagespunkten. metric:MOOD\n\nZuletzt aktualisiert"
metric:BLOOD_PRESSURE_SYS	"…regelmäßig nur eine von zwei geplanten Tagesdosen. metric:BLOOD_PRESSURE_SYS\n\nZuletzt aktualisiert"

The Zuletzt aktualisiert: … timestamp tail is the per-section advisor card footer — so the leak is appearing at the END of an advisor card recommendation text (not in the summary or primaryRecommendation, both of which are stripped).

Where the strip is missing

src/components/insights/recommendation-card.tsx:336:

<p className="text-sm leading-snug">{norm.text}</p>

norm.text (= rec.text) is rendered directly without stripChartTokens(). Every other AI-prose surface on the page — insight.summary (line 594), insight.primaryRecommendation (line 573), finding.label (line 622), finding.guideline (line 631) — runs through stripChartTokens(). The recommendation text is the one field the v1.4.19 A3 widening missed.

The model is following its own prompt (it embeds metric:WEIGHT etc. in the rec text expecting an inline mini-chart), but the rec-card's chart slot lives in the EXPANDED rationale block (line 226-234, only visible after the user clicks the chevron), so the token has nowhere to be consumed in the always-visible rec text — it just leaks.

Concrete fix

Two-line diff:

- <p className="text-sm leading-snug">{norm.text}</p>
+ <p className="text-sm leading-snug">{stripChartTokens(norm.text)}</p>

plus the import { stripChartTokens } from "@/lib/insights/chart-tokens"; at the top of recommendation-card.tsx.

Backstop test in the existing insights-polish.test.ts describe-block (v1.4.19 A3 — chart-token leak hardening) — add an it("RecommendationCard strips metric:* tokens from rec.text") case that mounts a rec with text: "increase fluids. metric:WEIGHT" and asserts the rendered DOM does NOT contain metric:WEIGHT.

Bonus translation defect uncovered en route

Marc reported "Metrik Mut" — the actual rendering on prod is the Health-Score sub-bar label, which surfaces "Mood" verbatim in German because messages/de.json:778 says:

"componentMood": "Mood",

instead of "Stimmung". The label is rendered by src/components/insights/health-score-card.tsx:65 via COMPONENT_LABEL_KEY.mood = "insights.healthScore.componentMood".

One-character fix. Worth bundling into the same Wave-4 patch so "the BP/Weight/Mood/Meds row" reads natively German.

4. Targets / Zielwerte page

Current state (probe screenshots)

/tmp/v1422-w1a-targets/desktop.png (1280×900) and /tmp/v1422-w1a-targets/pixel5.png (393×851).

The page is a 2-col (desktop) / 1-col (mobile) grid of identical <TargetCard> instances. Each card has:

• Card header: icon + label + small trend arrow (TrendingUp/Down/Minus)
• Big-number current value (with unit and 30-day mean line)
• Horizontal three-zone range bar (red / yellow / green) with a dot marker for the current value
• Status pill (e.g. "Optimal", "Im Zielbereich", "Moderat")
• External-link to the source guideline (WHO BMI / ESH 2023 / CDC/NCHS …)

Compared to /dashboard and /insights the page is visually static and reference-y:

• No 7-day delta. Only the 30-day mean.
• No sparkline / mini-chart in the card. The user sees the current value but no story of how they got there.
• No "Δ vs last month" callout (the comparison toggle pattern shipped in v1.4.16 phase B8 lives only on the dashboard).
• No grouping. BD-Zielbereich (50 % in target) sits next to BMI (Normal) sits next to Einnahmetreue (64 %) — three completely different metrics styled identically.
• Status pills are passive — they label the current bucket but don't invite action ("So what do I do about this?").

The "ungelebt" feeling is real: every card is a flat reference card with no momentum, no comparison, no "your trajectory" angle. It looks like a guideline cheat-sheet, not a living health log.

Three upgrade directions (pick one for Wave 4)

A) Sparkline-per-card + "Δ vs last month" callout

Each <TargetCard> grows a 30-day mini-sparkline beneath the range bar (same <HealthChart mini> already used in the InlineCharts of /insights). A second-line "Δ −2.3 kg vs last month" caption mirrors the dashboard's comparison toggle so the user sees movement at a glance. Status pill stays.

Pro: reuses existing components (HealthChart mini, comparison delta, TrendCard's metric-sentiment colour rules). Each card becomes a "current state + recent journey" block.

Con: doubles card height. 12 cards × +120 px is +1440 px scroll on mobile. Need a "compact mode" toggle.

B) Status-row + drill-down — collapse cards into a scannable list

Replace the grid with a vertical list. Each row: icon + label + current value + status pill + small trend arrow + chevron to expand. Expansion reveals the range bar + 30-day chart + source link.

Pro: 12 metrics fit on one screen. Glance answer first, detail on demand. Mirrors the iOS Health pattern users know.

Con: Bigger redesign — at-a-glance loses the range-bar visualisation that gives the page its "where am I in the band?" answer. Might be too clinical.

C) Storyline framing — a short headline + 3-target hero, then everything else

The page opens with a one-sentence summary "Du bist in 8 von 12 Zielbereichen.", then surfaces the 3 most-out-of-target metrics as hero cards with bigger range bars + a "What changed?" callout, with the remaining 9 collapsed below as a compact summary strip.

Pro: solves the "ungelebt" problem with a narrative. Targets become a state-of-your-health page, not a reference book. The hero strip ties back to the <HeroStrip> pattern shipped on /insights for visual consistency.

Con: requires server-side ranking ("which targets are most out of band?"), so this is a Phase-2 redesign. Out of scope for a polishing marathon. Defer to v1.5.

Recommendation: A (sparkline + Δ-vs-last-month)

Lowest effort, highest "the page now feels alive" payoff, reuses existing components. The mobile-height cost is mitigated by collapsing the range bar's surrounding meta into a single row. Direction B is a v1.5-scope rethink; direction C is a v1.6 product call.

For Wave 4 the lift is roughly:

• Add 30dPoints?: number[] to the TargetData API shape (server already has the data — /api/insights/targets reads from the same Prisma rows the dashboard does).
• Mount <HealthChart types={[target.type]} mini /> inside <TargetCard> between the range bar and the classification row.
• Pipe the comparison delta from the existing compare server route (used by the dashboard) into the page query, render it as a <TileCompareDelta> row above the range bar.

5. v1.4.21 backlog inventory

Source: .planning/v1421-backlog.md (301 lines). Items grouped by complexity bucket below.

Quick wins (apply en masse in Wave 4 D — ≤5 lines diff each)

• Sec-L-1 — coachChatRequestSchema.conversationId regex tightening to ^c[a-z0-9]{24}$ or [A-Za-z0-9_-]{8,64}. One-line schema change.
• Sr-LOW-1 — drop redundant @@index on coach_usage; @@unique already builds it. One-line schema removal.
• Code-MED-04 — reuse buildWeightRangeFromHeight() for weightTrendAlignment's ±2 kg target band. Single import + call swap.
• Code-MED-07 — drop the dead confidence prop from <TrendsRow> call site. One-line removal.
• Design-L1 — drop the generic "personal-baseline" hero copy line. One-line removal.
• Design-L5 — append … suffix when storyboard 24-char truncation fires. Three-line addition in src/components/insights/storyboard-row.tsx.
• Design-L6 — replace German Umschalt+Enter with literal Shift+Enter to match the rest of the app. One i18n key change.
• Code-LOW-01 — wrap <HeroStrip> greeting + relativetime in useMemo. Two-line change.
• Code-MED-09 — wrap storyboard payload through storyboardAnnotationsSchema.safeParse. Pattern-copy from dailyBriefing lift; ~5 lines.
• Sr-MED-1 — add 3 cascade-delete test rows for coachConversation / coachMessage / coachUsage. 5 lines.
• Sr-LOW-4 — tighten weeklyReport.weekISO regex to validate weeks 01-53. One regex change.
• S-06 — drop dead historyRail / sourcesRail / composer slot props on <CoachDrawer>. ~10 lines deletion.
• S-07 — drop the never-overridden inputId prop on <CoachInput>. One-line removal.

Wave-4 quick-wins count: 13 items, total estimated diff ≤ ~80 lines across 11 files. Recommend folding all 13 into Wave-4 D.

Medium (one new helper / one component refactor)

• Code-MED-02 — bandFromInterval "high" label on near-zero r — rename chip to "tight CI" / "wide CI" or demote when CI straddles zero. Touches the helper + every chip caller.
• Code-MED-03 — Pearson p-value normal-approx error at low df. Either pull in incomplete-beta (~30 LOC) or raise df gate to 20. Decision call.
• Code-MED-05 — bpInTargetPct "vs last week" delta is held constant. Recompute prior-snapshot rate or relabel the line.
• Code-MED-06 — WeeklyReportView print spinner timer reset bug. Pin via ref.
• Code-MED-08 — provenanceFromJson cast — validate against WINDOW_KEYS allow-list.
• Code-LOW-03 — Promote nested <Sheet> portals on mobile rail trays to siblings.
• Code-LOW-05 — add bpInTargetRate constant-test gap test.
• Code-LOW-06 — streamRefusal budget meter accounting.
• Sec-M-2 — call checkRateLimit() before refusal scanning OR bump messageCount inside streamRefusal().
• Sec-M-3 — extend INJECTION_PATTERNS with synonyms + zero-width strip pass. Maintenance pattern, ~30 LOC.
• Sec-M-4 — wrap appendMessage(assistant) + recordSpend() in Prisma $transaction.
• Sec-L-3 — skip updatedAt bump for refusal appends OR hide refusal-only conversations.
• Sec-U-1 — Vitest case smoke-testing request.clone() against Next.js 16.
• Sr-MED-2 — lift weeklyReport + storyboardAnnotations through safeParse.
• Sr-MED-3 — wrap handleChatRequest in enforceRateLimit().
• Sr-MED-6 — offer withIdempotency(handler, { skipWhen }) for conditional opt-out.
• Sr-LOW-2 — migrate metricSourceJson from String to Jsonb.
• Sr-LOW-3 — coach off-topic detector locale-mixed tightening.
• S-08 — re-read cloneForCheck dead variable.
• S-09 — compress health-score-card delta-arrow && blocks into a deltaIcon lookup. Borderline.
• Design-M1 — Hero strip greeting clock → Berlin via Intl.DateTimeFormat.
• Design-M3 — Health Score panel mobile order inversion.
• Design-M5 — Coach drawer focus trap — reparent rail-tray Portal container to SheetContent.
• Design-M6 — Tone-bar visual clip at rounded corner.
• Design-M8 — "n=N" reads as engineering output — use {count} samples / {count} Werte.
• Design-L2 — Health Score sub-bars per-component band.

Wave-4 medium count: 26 items. Recommend Wave-4 picks the BP-tile fix (Code-MED-05), the Sec-M-2/3/4 cluster (small but security-adjacent), the safeParse lifts (Sr-MED-2 + Code-MED-08 + Code-MED-09), and the 2-3 design polishes that ride along (Design-M1 / M6 / M8). Defer the rest to v1.4.23.

Heavy (cross-cutting, schema, or larger refactor)

• Sr-HIGH-2 — Consolidate Pearson + linear regression duplicated maths layer into correlation.ts + regression.ts. v1.4.23.
• Sr-HIGH-4 — <CoachDrawer key={prefill}> weaponising React keys for state reset. Make prefill fully-controlled. v1.4.23.
• Sr-MED-4 — Wire provider runner to true streaming with onToken callback. Unlocks "stop generating" as a real feature. v1.5.
• Sr-MED-5 — medication_schedules.days_of_week schema drift. Either land the column with backfill or drop the field. v1.4.23 (schema change with downstream impact).
• Code-MED-01 — Streaming auto-scroll deps brittleness fix. Touches MessageThread effect and a new turn-counter hook. v1.4.23.

Wave-4 heavy count: 5 items, all deferred.

Already obsolete (resolved en passant or no longer relevant)

• Sec-U-2 — enforceBudget race window. Acceptable in practice; per-day cap. Annotation-only suggestion. Promote to "not a defect" and close.
• Sr-MED-6 carry-over — apiHandler-vs-withIdempotency stacking is "largely moot now that the chat route doesn't wrap in idempotency". The carry-over note documents that — close as obsolete.
• The Docs-site MED + LOW items (versions stale, model count, schedule timezone, hashing wording) — these are cross-repo to healthlog-docs. Out of scope for v1.4.22 HealthLog work; belong in the docs repo's own next sweep. Move to a separate file docs-followup.md or close as out-of-repo.
• F5 README badges, FUNDING.yml, .gitignore audit, repo-root tidy — cosmetic only. The README badge add is useful but not v1.4.22-blocking. Push to a hygiene PR.
• CLAUDE.md / AGENTS.md filename rename, source-comment "Marc" sweep, DE+EN bilingual CHANGELOG normalisation — defer to a hygiene PR series in v1.5. The filename rename in particular is structural and would affect every IDE / CI tool that special-cases these filenames; a one-line v1.4.22 patch is the wrong vehicle.

Obsolete count: 7 items. Mark closed in the v1.4.21 backlog, move docs items to docs-followup.md, and rebase v1.4.22 backlog to start with the 13 quick-wins + 8-10 medium picks above.

---

Summary numbers

• Quick-wins to fold into Wave-4 D: 13 items, ~80 LOC total
• Medium picks to fold into Wave-4 (selectively): 8-10 of 26
• Heavy items deferred to v1.4.23 / v1.5: 5
• Obsolete to close: 7

Probe artifacts:

• /tmp/v1422-w1a/findings.json + 4 viewport screenshots
• /tmp/v1422-w1a-insights/findings.json + 3 locale screenshots
• /tmp/v1422-w1a-targets/desktop.png + pixel5.png