fix(supply-squeeze): per-SKU paired price comparison eliminates catalog-composition bias

The 30d-vs-60d price momentum aggregated AVG/median across whatever SKUs
happened to be in a speed/form-factor bucket each period. New expensive SKUs
entering the catalog (NVIDIA switches at 30k USD, AOC cables) faked huge jumps
— 400G OSFP showed +151% when matched-SKU reality was 0%.

Now: compute per-transceiver median price in each period, keep only SKUs present
in BOTH periods (>=2 obs each), report the median of per-SKU pct deltas. Also
excludes non-transceiver form factors, AOC/DAC cables, switch SKUs, price>15k,
and anomalous observations. Result: 400G OSFP +151%->0%, signals 21->8, and the
ones that remain (NVIDIA MFA7U10 +84% same-SKU) are genuine price moves.

packages/api/src/routes/procurement.ts

@@ -787,16 +787,43 @@ procurementRouter.get("/supply-squeeze", async (_req: Request, res: Response) =>
     const [priceSignals, aiDemand, hypeData, stockData] = await Promise.all([
       // Price momentum: 30d vs 60d avg by speed/form_factor
       pool.query(`
+        -- Per-SKU paired comparison: only transceivers present in BOTH periods.
+        -- This eliminates catalog-composition bias (new expensive SKUs entering a
+        -- speed/form-factor bucket would otherwise fake a huge price jump).
+        WITH per_sku AS (
           SELECT
-          t.speed_gbps, t.form_factor,
+            t.id, t.speed_gbps, t.form_factor,
-          ROUND(AVG(po.price) FILTER (WHERE po.time >= NOW() - INTERVAL '30 days')::numeric,2) AS avg_30d,
+            PERCENTILE_CONT(0.5) WITHIN GROUP (ORDER BY po.price)
-          ROUND(AVG(po.price) FILTER (WHERE po.time >= NOW() - INTERVAL '60 days' AND po.time < NOW() - INTERVAL '30 days')::numeric,2) AS avg_prior_30d,
+              FILTER (WHERE po.time >= NOW() - INTERVAL '30 days') AS med_now,
-          COUNT(*) FILTER (WHERE po.time >= NOW() - INTERVAL '30 days') AS obs_30d
+            PERCENTILE_CONT(0.5) WITHIN GROUP (ORDER BY po.price)
+              FILTER (WHERE po.time >= NOW() - INTERVAL '60 days' AND po.time < NOW() - INTERVAL '30 days') AS med_prior,
+            COUNT(*) FILTER (WHERE po.time >= NOW() - INTERVAL '30 days') AS obs_now,
+            COUNT(*) FILTER (WHERE po.time >= NOW() - INTERVAL '60 days' AND po.time < NOW() - INTERVAL '30 days') AS obs_prior
           FROM price_observations po
           JOIN transceivers t ON t.id = po.transceiver_id
           WHERE po.price > 5 AND po.currency = 'USD'
-        GROUP BY t.speed_gbps, t.form_factor
+            AND COALESCE(po.is_anomalous, false) = false
-        HAVING COUNT(*) FILTER (WHERE po.time >= NOW() - INTERVAL '30 days') >= 3
+            AND t.form_factor IN ('SFP','SFP+','SFP28','SFP56','QSFP+','QSFP28','QSFP56','QSFP-DD','QSFP-DD800','OSFP','OSFP-XD','XFP','CFP','CFP2','CFP4','CDFP','DSFP')
+            AND po.price < 15000
+            AND t.part_number NOT ILIKE '%AOC%'
+            AND t.part_number NOT ILIKE '%-DAC-%'
+            AND (t.standard_name IS NULL OR (t.standard_name NOT ILIKE '%Switch%' AND t.standard_name NOT ILIKE '%InfiniBand%'))
+          GROUP BY t.id, t.speed_gbps, t.form_factor
+        )
+        SELECT
+          speed_gbps, form_factor,
+          -- avg_30d / avg_prior_30d kept as column names for downstream compatibility,
+          -- but they now carry MEDIAN-of-matched-SKU prices (only SKUs in both periods)
+          PERCENTILE_CONT(0.5) WITHIN GROUP (ORDER BY med_now)   AS avg_30d,
+          PERCENTILE_CONT(0.5) WITHIN GROUP (ORDER BY med_prior) AS avg_prior_30d,
+          -- The real signal: median of per-SKU percentage deltas
+          ROUND(PERCENTILE_CONT(0.5) WITHIN GROUP (ORDER BY (med_now - med_prior) / NULLIF(med_prior,0) * 100)::numeric, 1) AS sku_median_delta_pct,
+          COUNT(*) AS obs_30d
+        FROM per_sku
+        WHERE med_now IS NOT NULL AND med_prior IS NOT NULL AND med_prior > 0
+          AND obs_now >= 2 AND obs_prior >= 2
+        GROUP BY speed_gbps, form_factor
+        HAVING COUNT(*) >= 3
       `),
       // AI cluster demand by speed tier
       pool.query(`
@@ -835,7 +862,7 @@ procurementRouter.get("/supply-squeeze", async (_req: Request, res: Response) =>
       `).catch(() => ({ rows: [] })),
     ]);
 
-    type PriceRow = { speed_gbps: string; form_factor: string; avg_30d: string; avg_prior_30d: string; obs_30d: string };
+    type PriceRow = { speed_gbps: string; form_factor: string; avg_30d: string; avg_prior_30d: string; obs_30d: string; sku_median_delta_pct: string | null };
     type HypeRow = { technology: string; hype_phase: string; hype_score: string };
     type AiRow = { speed_tier: string; total_tx: string; cluster_count: string };
     type StockRow = { speed_gbps: string; form_factor: string; out_of_stock: string; in_stock: string; total_obs: string };
@@ -861,9 +888,12 @@ procurementRouter.get("/supply-squeeze", async (_req: Request, res: Response) =>
     const signals = (priceSignals.rows as PriceRow[])
       .map((r) => {
         const speed = parseFloat(r.speed_gbps);
-        const priceUp = r.avg_30d && r.avg_prior_30d
+        // Prefer the per-SKU median delta (composition-bias-free); fall back to aggregate
+        const priceUp = r.sku_median_delta_pct != null
+          ? parseFloat(r.sku_median_delta_pct)
+          : (r.avg_30d && r.avg_prior_30d
               ? ((parseFloat(r.avg_30d) - parseFloat(r.avg_prior_30d)) / parseFloat(r.avg_prior_30d)) * 100
-          : 0;
+              : 0);
         const hype = speedToHype.get(speed);
         const ai  = aiBySpeed.get(speed);
         const stock = stockByKey.get(`${r.speed_gbps}:${r.form_factor}`);