mirror of
https://github.com/Kpa-clawbot/meshcore-analyzer.git
synced 2026-04-27 14:55:15 +00:00
Kpa-clawbot
a0fddb50aa
Closes #789. ## The two bugs 1. **Severity from stale median.** `classifySkew(absMedian)` used the all-time `MedianSkewSec` over every advert ever recorded for the node. A repeater that was off for hours and then GPS-corrected stayed pinned to `absurd` because hundreds of historical bad samples poisoned the median. Reporter's case: `medianSkewSec: -59,063,561.8` while `lastSkewSec: -0.8` — current health was perfect, dashboard said catastrophic. 2. **Drift from a single correction jump.** Drift used OLS over every `(ts, skew)` pair, with no outlier rejection. A single GPS-correction event (skew jumps millions of seconds in ~30s) dominated the regression and produced `+1,793,549.9 s/day` — physically nonsense; the existing `maxReasonableDriftPerDay` cap then zeroed it (better than absurd, but still useless). ## The two fixes 1. **Recent-window severity.** New field `recentMedianSkewSec` = median over the last `N=5` samples or last `1h`, whichever is narrower (more current view). Severity now derives from `abs(recentMedianSkewSec)`. `MeanSkewSec`, `MedianSkewSec`, `LastSkewSec` are preserved unchanged so the frontend, fleet view, and any external consumers continue to work. 2. **Theil-Sen drift with outlier filter.** Drift now uses the Theil-Sen estimator (median of all pairwise slopes — textbook robust regression, ~29% breakdown point) on a series pre-filtered to drop samples whose skew jumps more than `maxPlausibleSkewJumpSec = 60s` from the previous accepted point. Real µC drift is fractions of a second per advert; clock corrections fall well outside. Capped at `theilSenMaxPoints = 200` (most-recent) so O(n²) stays bounded for chatty nodes. ## What stays the same - Epoch-0 / out-of-range advert filter (PR #769). - `minDriftSamples = 5` floor. - `maxReasonableDriftPerDay = 86400` hard backstop. - API shape: only additions (`recentMedianSkewSec`); no fields removed or renamed. ## Tests All in `cmd/server/clock_skew_test.go`: - `TestSeverityUsesRecentNotMedian` — 100 bad samples (-60s) + 5 good (-1s) → severity = `ok`, historical median still huge. - `TestDriftRejectsCorrectionJump` — 30 min of clean linear drift + one 1000s jump → drift small (~12 s/day). - `TestTheilSenMatchesOLSWhenClean` — clean linear data, Theil-Sen within ~1% of OLS. - `TestReporterScenario_789` — exact reproducer: 1662 samples, 1657 @ -683 days then 5 @ -1s → severity `ok`, `recentMedianSkewSec ≈ 0`, drift bounded; legacy `medianSkewSec` preserved as historical context. `go test ./... -count=1` (cmd/server) and `node test-frontend-helpers.js` both pass. --------- Co-authored-by: clawbot <bot@corescope.local> Co-authored-by: you <you@example.com>