[adminopenclaw8-sketch]

Relates to #304

Points 1, 2 and 6 of the issue landed in #351 and #352. This PR implements the
remaining points 3 (status evidence), 4 (SNR attribution) and
5 (traffic metrics) on the My Mesh cards and the home health panel, and
builds on the observer identity #572 (#262) just added.

Plan

1. Check what the node-health payload actually proves before touching the UI
   (AGENTS.md rule 4): stats.lastAdvert is strictly the node's own ADVERT,
   stats.lastHeard is that advert or identity-safe relay evidence
   (whichever is newer), stats.avgSnr is already the mean over direct
   receptions since Node "Heard By" card: decide whether to show direct-RF observers (defer) #76, and activity24h is the rolling window from My Mesh: accurate 24h node activity and compact repeater observer cards #351.
2. Put the labelling rules in one shared, pure helper
   (public/my-mesh-activity.js, the shape My Mesh: show observer identity, status and traffic separately from mesh role #262 established) so the cards and
   the health panel cannot drift apart, and so every rule is unit-testable.
3. Add the only missing datum — the previous 24-hour baseline — to the window
   the card already fetches, never as a new request.
4. Keep the My Mesh: accurate 24h node activity and compact repeater observer cards #351 compactness guarantee: new explanation goes into a native
   <details>, and the metric grid loses a tile.

What changed

3. Status evidence

• Last own advert and Last attributed relay activity are separate
  rows in a new Status evidence disclosure on each card. The relay row is
  only filled where the evidence supports it: the payload exposes one combined
  lastHeard, so relay activity is reported exactly when it is newer than
  the node's own advert, and otherwise says none newer than the last own advert rather than repeating the advert's time.
• The badge names its evidence. When the node's own advert alone supports
  the badge the dot stays filled and green. When the only evidence recent
  enough is traffic attributed to the node by a short wire-hash prefix, the
  badge becomes Active · relay-attributed, the dot is hollow and the card's
  accent drops to the provisional colour — an ambiguous match never produces
  a confident green. The disclosure explains that a prefix unique among
  known relays can still belong to an unknown one, and that a recent advert
  is not proof of continuous uptime.
• No new thresholds. getHealthStatus / getHealthLastHeard (roles.js,
  fix(health): apply role thresholds without weakening activity evidence #539) remain the single source of truth and are injectable; the helper
  defines none and the test asserts it.

4. SNR

• The Excellent / Good / Marginal / Poor verdict and its colour are removed
  from both the card and the health panel. Until Node "Heard By" card: decide whether to show direct-RF observers (defer) #76's attribution decision
  is made, the card explains the metric instead of grading the repeater.
• The label states the sample scope the mean covers (Direct SNR · n=42), and
  the value is only shown when at least one direct reception proves it belongs
  to this node — a mean with no direct attribution behind it would describe
  whichever node the observer actually heard.
• stats.avgSnr itself is untouched: no shared health API metric is
  redefined.

5. Traffic

• Packets today (a UTC calendar-day counter) becomes
  Transmissions · 24 h, the rolling window's own count, with the
  population and window in the tooltip and as a Counting window row in the
  disclosure (a title is not reachable by touch).
• Incomplete coverage renders as a lower bound (≥ 35) and an unknown window
  as an em dash — never as a measured zero.
• activity24h gains previous (windowStart, windowEnd, count,
  complete): the equally long window right before the rolling one. The
  comparison is shown only when both windows are proven complete, a
  missing history is reported as no comparable 24 h history, and a zero
  baseline yields an absolute delta, never a division.
• Avg hops leaves the headline grid for the disclosure as
  Avg observed path length, explained as the mean hop count of the
  observed paths of the node's transmissions — not a distance and not health.
  Nothing labels node-associated traffic as forwarded traffic.

Cost and compactness

• No new request. Everything comes from the node-health call the card
  already makes (/api/nodes/{key}/health?include=advertIntervals); the five
  api( call sites in home.js are unchanged and pinned by a test.
• Server side, previous is filled by the same single pass over the
  node's index that already fills the hourly buckets. The only added cost is
  the RFC3339 parse for rows 24–48 h old that the string pre-filter used to
  reject — the same order as the per-transmission work GetNodeHealth already
  does for every row (it calls heard.add on each). No new scan, no new lock,
  no new allocation per row.
• The card stays compact: the metric grid is now three tiles on one row at
  every width (it was four over two rows below 600 px) and the explanation
  is a closed <details>. Measured on the 51-observer repeater card:
  1280 px 337.8 → 343.8, 390/375 px 349.1 → 329.1 — still inside the < 350
  bound My Mesh: accurate 24h node activity and compact repeater observer cards #351 pinned, and lower than before on mobile.

Testing

Red before, green after for every point, each with at least one mutant; see
the report comment for the full table. Locally: go test ./... in
cmd/server and cmd/ingestor, sh test-all.sh (271 files),
node test-frontend-helpers.js, and the affected Playwright suites against a
local Go server on the committed e2e-fixture.db prepared as in CI.

[adminopenclaw8-sketch]

Rapport — CS-Macmini PR#586 #304 — head a6b8c45

Status: All three remaining #304 points implemented; every point has a test that is red before and green after plus at least one killed mutant; all 17 CI jobs pass.

Requirement coverage

#	Requirement	Implementation	Test (red before → green after)	Mutant (killed)
3a	Last own advert and Last attributed relay activity shown separately, only where the evidence supports it	statusEvidence() in public/my-mesh-activity.js; two rows in the card's Status evidence disclosure. Relay activity is reported exactly when lastHeard is newer than lastAdvert, otherwise none newer than the last own advert	[T] test-issue-304-my-mesh-metrics.js — "the two evidence rows are labelled separately and never merged", "an advert that alone supports the badge stays confident even with newer relay traffic"; [T] test-issue-304-my-mesh-e2e.js rows 0/1/2	M3 relayAt = heardAt (relay row echoes the advert time) → killed
3b	Explain what drives the status badge	ev.explanation per source (advert / relay / none), rendered in the disclosure; badge text carries the source	[T] metrics test — explanation asserted per source, incl. "not proof of continuous uptime"; [T] E2E .mnc-evidence-note per card	M1 confident: true (badge never names attributed evidence) → killed
3c	Ambiguous short-prefix relay matches must not give a confident green status	Attributed status → hollow dot (ph-circle), status-warn class, Active · relay-attributed, and .my-node-card.healthy.evidence-relay drops the green accent	[T] metrics test — "relay-only evidence never produces a confident green status", "a node with no own advert at all is relay-attributed, never confident"; [T] E2E — dot class, glyph href, card class and computed border-left-color vs the resolved --status-green	M2 cls: statusCls (green dot kept) → killed; [T] E2E mutant E2 (remove the accent rule) → killed
3d	Reuse the existing liveness rules, no new thresholds	getHealthStatus / getHealthLastHeard (roles.js, #539) injected, defaulted from the window globals; the helper defines none	[T] metrics test — "the role thresholds come from roles.js, with no home-page copy" (infra 90 min healthy vs client degraded, plus a source scan for a threshold definition and an injected classifier); [T] test-health-status.js executes the real production decision blocks	M (helper declaring its own HEALTH_THRESHOLDS) → asserted by source scan
4a	Do not present the representative observation's SNR as the repeater's own signal quality	Verdict words and verdict colour removed from the card and the health panel; tooltip states the value is measured at the observer and is not a quality or health grade	[T] metrics test — "the direct SNR mean carries no quality verdict and no verdict colour"; [T] static pins on home.js (snrLabel/snrColor gone, no verdict words); [T] E2E — metric text has no verdict, .mnc-metrics .mnc-val[style] count 0	[T] E2E mutant E4 (re-add an inline verdict colour) → killed
4b	Explain the metric / sample scope; attribution must support the claim	Label Direct SNR · n=<receptions>; value only when at least one direct observer proves it (#76)	[T] metrics test — "the SNR label states the sample scope", "an SNR with no direct attribution behind it is not presented at all", "a non-numeric packet count never fabricates a sample size"; [T] E2E — card 0 n=30, card 1 no n=, card 2 em dash	M4 available = value !== null (SNR shown with no direct attribution) → killed
4c	Do not redefine shared health API metrics	stats.avgSnr untouched server-side; only the client label changed	[A] cmd/server diff touches no SNR code; GetNodeHealth SNR path unchanged	—
5a	Clearly labelled rolling 24 h count with population and window, replacing "Packets today"	traffic24h() sums the rolling window's own buckets; label Transmissions · 24 h; population + window in the tooltip and as a Counting window row (a title is not touch-reachable)	[T] metrics test — "the traffic metric is a labelled rolling 24h count, not 'packets today'", "the window text states both bounds and the coverage caveat"; [T] E2E labels/values; [T] static pin: no Packets today left in home.js	—
5b	Missing/incomplete coverage ≠ zero	Incomplete window renders ≥ N; an unknown/malformed window renders an em dash	[T] metrics test — "an incomplete window is a lower bound, never a measured total", "a missing or malformed window is unavailable, not zero", "a genuine quiet window with proven coverage is an exact zero"; [T] E2E cards 1 and 2	M5 exact = true (lower bound shown as a total) → killed
5c	Previous-24 h comparison only with comparable history; no division by 0; missing history ≠ 0	activity24h.previous added server-side (same single pass); compare24h() requires both windows complete and a genuine integer baseline; a zero baseline yields an absolute delta, never a percentage	[T] metrics test — "the comparison needs two complete, like-for-like windows", "an unknown previous window is never treated as zero traffic", "a zero baseline produces no percentage and never divides by zero"; [T] cmd/server/node_activity_304_test.go × 3; [T] E2E — chip on card 0, absent on 1 and 2	M6 baseline = prev.count ?? 0 → killed; M7 (skip the zero-baseline branch) → killed; G1 Previous.Complete = a.Complete → killed; G4 previous window 12 h instead of 24 h → killed
5d	Move/explain Avg hops	Out of the headline grid, into the disclosure as Avg observed path length with its route/sample meaning and an explicit denial of the distance reading	[T] metrics test — "avg hops is relabelled as an observed path length, with its sample scope"; [T] E2E — avg hops absent from .mnc-metrics, row present in the disclosure (10 hops / —)	M8 label reverted to Avg hops → killed
5e	Do not label node-associated traffic as forwarded traffic	Tooltip states the population (own / addressed to it / observed path names it) and that it is not a forwarding count	[T] metrics test asserts the tooltip never sells the count as forwarded traffic	—
—	API shapes checked before UI (AGENTS.md rule 4)	[A] NodeHealthStats (cmd/server/node_activity.go), heardByAccumulator.avgSnr (heard_by.go), NodeActivity24h (node_activity_24h.go), openapi.go read before writing client code	—	—
—	No per-card API call	Everything comes from the node-health request the card already makes	[T] metrics test — "the card adds no API request of its own" (the five api( call sites are pinned)	—
—	Unit tests for labels and status logic, E2E incl. mobile 390 px, CSS variables	New test-issue-304-my-mesh-metrics.js (26 cases); test-issue-304-my-mesh-e2e.js runs at 1280 px and 375 px with touch	[T] CSS block scanned for hardcoded colors; [K] node scripts/check-css-vars.js clean	[T] E2E mutant E3 (disclosure open by default) → killed; E1 (summary click navigates) → killed

Server-side change

activity24h gains previous {windowStart, windowEnd, count, complete} — the equally long window before the rolling one, filled by the same single pass over the node's index. The only added cost is the RFC3339 parse for rows 24–48 h old that the string pre-filter used to reject, which is the same order as the per-transmission work GetNodeHealth already performs on every row. [A] No new scan, no new lock, no new allocation per row; cmd/server stays read-only and no map[string]interface{} was added outside tests.

Compactness (#351's bound preserved)

[K] Measured on the 51-observer repeater card: 1280 px 337.8 → 343.8 px, 390/375 px 349.1 → 329.1 px. Both stay inside the < 350 bound #351 pinned, and mobile is 20 px shorter than before: the metric grid is now three tiles on one row at every width (it was four over two rows below 600 px) and the explanation lives in a closed <details>.

CI (run 40172746937, head a6b8c45)

Job	Result
✅ Go Build & Test	pass
✅ Go Server Tests (1/3)	pass
✅ Go Server Tests (2/3)	pass
✅ Go Server Tests (3/3)	pass
✅ Go Ingestor Tests	pass
✅ Go Libraries & CLI	pass
✅ JavaScript Unit Tests & Lint	pass
✅ Script & Container Unit Checks	pass
🎭 Playwright E2E Tests (1/3)	pass
🎭 Playwright E2E Tests (2/3)	pass
🎭 Playwright E2E Tests (3/3)	pass
📊 Verify E2E Shards & Coverage	pass
📋 Verify Go Test Assignments	pass
🧱 Validate Docker Image	pass
🏗️ Build & Publish Docker Image	pass
📦 Release Artifacts / 🚀 Deploy Staging / 📝 Publish Badges	skipped (fork guards)

The first push failed 🧹 Frontend lint (eslint no-undef) only: a new window-scoped helper needs a .eslintrc.json globals entry, exactly as MyMeshObservers did in #262. Fixed in a6b8c456; no other job failed and the known-flaky #271 suite did not fire.

Local before pushing: go test ./... in cmd/server and cmd/ingestor, sh test-all.sh (271 files), node test-frontend-helpers.js (725 cases), and the affected Playwright suites (#304, #262, #2027, #76, home coverage, health status, the 157-test core suite) against a local Go server on the committed e2e-fixture.db, prepared exactly as CI does.

Checks

[K] scripts/check-xss-sinks.sh --diff origin/master → exit 0, no finding. [K] node scripts/check-css-vars.js → 0 undefined. [K] Fork guards unchanged: 9 in deploy.yml, 1 in release-fast-path.yml. [A] cmd/server remains read-only; no new map[string]interface{} outside tests; no hardcoded colors (the new CSS block is scanned by the unit test). [A] No staging or production access, no server API key used, no upstream write.

Adjacent tests updated (behaviour change, not weakened assertions)

• test-health-status.js — the two home decision blocks it executes now go through the shared helper, so the extraction markers moved and the helper is loaded into the same VM context. The blocks executed are still the real production code.
• test-health-status-e2e.js — the health banner's aria-label is now the user-facing badge text instead of the raw status token. The expectation maps status → label and adds an assertion: the confirmed-relay and legacy-directory fixtures (neither has an own advert) must render the attributed glyph, the others the confirmed one.
• test-issue-262-mymesh-observer-e2e.js — its fixtures gained lastAdvert and a complete rolling window, so its "node activity stays separate from observer reporting" assertion now reads the rolling-24 h metric instead of the removed Packets today.
• test-issue-2027-my-mesh-node-page-e2e.js, test-issue-262-mymesh-observer-e2e.js — load the new helper alongside my-mesh-observers.js, as index.html does.

Remainder

• Point 4 stays deliberately descriptive: the card explains the direct-SNR metric rather than grading the repeater, pending the attribution decision in Node "Heard By" card: decide whether to show direct-RF observers (defer) #76. Nothing here pre-empts that decision.
• The direct-SNR sample size counts receptions across all direct observers; a per-observer breakdown stays on the Heard By tables.
• getHealthLastHeard's legacy directory fallback (a response without lastHeard) now reads as relay-attributed, since no own advert backs it. That is honest for the fallback, but it means an old cached payload shows the provisional badge; the live API always sends lastHeard.
• Not filed: a follow-up could expose the relay evidence timestamp separately from lastHeard, which would let the card show relay activity that is older than the last own advert. That needs a second pass over the relay candidates and was left out on cost grounds.