fix: catch non-zero exit in local train, return JSON instead of 500

- sh -lc replaced with bash to avoid dash/profile.d syntax errors
- runCommand errors now caught in local provider path
- stdout/stderr extracted from error object and returned as JSON
- No more HTTP 500 on script failure

packages/api/src/routes/selflearning.ts

@@ -165,8 +165,23 @@ selflearningRouter.post("/train", async (req: Request, res: Response) => {
       res.status(409).json({ success: false, error: "Local training command is not configured.", suggestion: "Set TIP_LOCAL_TRAIN_COMMAND; the lane name is appended automatically." });
       return;
     }
-    const out = await runCommand("sh", ["-lc", `${command} ${lane}`], 12 * 60 * 60 * 1000);
-    res.json({ success: true, provider, lane, stdout: out.stdout.slice(-6000), stderr: out.stderr.slice(-4000) });
+    try {
+      const out = await runCommand("bash", [command, lane], 12 * 60 * 60 * 1000);
+      res.json({ success: true, provider, lane, stdout: out.stdout.slice(-6000), stderr: out.stderr.slice(-4000) });
+    } catch (err: unknown) {
+      // execFileAsync throws on non-zero exit — stdout/stderr are still on the error object
+      const e = err as { stdout?: string; stderr?: string; message?: string };
+      const stdout = (e.stdout ?? "").slice(-6000);
+      const stderr = (e.stderr ?? "").slice(-4000);
+      res.json({
+        success: false,
+        provider,
+        lane,
+        error: e.message ?? String(err),
+        stdout,
+        stderr,
+      });
+    }
     return;
   }
   const endpoint = process.env.TIP_RUNPOD_ENDPOINT_ID || process.env.RUNPOD_ENDPOINT_ID;