Document real-world peak detector validation

PROJECT_KNOWLEDGE.md

@@ -26,7 +26,8 @@ after a value falls below `drop_ratio * current_peak`, it emits that maximum
 only after actual subsequent below-threshold observations support a continuous
 `hold_seconds` interval. `min_peak` is a generic detector setting, along with
 `drop_ratio` (0 < ratio < 1), `hold_seconds` (>= 0), and an explicit positive
-`max_sample_gap_seconds`. Elapsed time, never a sample count, determines
+`max_sample_gap_seconds`. They are selected per signal, rather than being
+globally fixed process constants. Elapsed time, never a sample count, determines
 confirmation. An interval longer than the configured maximum sample gap
 restarts a reset candidate, so a timestamp gap is not
 continuous-below evidence. Reset values can be negative rather than zero,
@@ -35,10 +36,25 @@ The maximum gap is required rather than inferred or defaulted, making each
 source's continuity assumption explicit.
 
 K7 roll length, variable `bf81c547-dccf-4709-aee2-0f78366d1dfc` (m), is the
-first real-world validation case. The tested capture is continuous on a
-10-second grid through its latest available sample; its shorter-than-requested
-result was caused by an end time queried in the future, not observed time-series
-gaps. The raw capture remains ignored and deterministic tests use synthetic
+clean reference validation case. With `min_peak=10`, `drop_ratio=0.75`,
+`hold_seconds=60`, and `max_sample_gap_seconds=20`, its continuous 10-second
+capture produced exactly four plausible peaks of approximately 49.98 m. Its
+shorter-than-requested result was caused by an end time queried in the future,
+not observed time-series gaps.
+
+Bento 2 roll weight is the second, more realistic robustness validation case:
+machine `6201f3de-b940-45d7-930c-1481c5ae9e1d`, variable
+`059c0015-9aff-4270-a4ce-f21b83f99378` (kg). Its observed cadence is 10
+seconds, but it has long plateaus, reset values roughly 300–470 kg, and
+materially varying peaks. With `min_peak=800`, `drop_ratio=0.75`,
+`hold_seconds=60`, and `max_sample_gap_seconds=20`, the extended capture
+yielded 13 plausible completed peaks: 936, 919, 935, 930, 930, 920, 909,
+1195, 921, 930, 939, 912, and 914 kg (from 00:46:50Z through 03:06:30Z on
+2026-09-04). The approximately 1195 kg peak is retained as a real
+process/mechanical condition, not special-cased as a defect; it is emitted
+only after its following reset has enough below-threshold observations. A
+reset is not expected to approach zero, and peak values are not assumed to be
+constant. Raw captures remain ignored and deterministic tests use synthetic
 fixtures only.
 
 ## Central domain context